6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciObjArray.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/opcodes.hpp"
45 #include "opto/parse.hpp"
46 #include "opto/reachability.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subtypenode.hpp"
50 #include "opto/type.hpp"
51 #include "runtime/deoptimization.hpp"
52 #include "runtime/sharedRuntime.hpp"
53 #include "utilities/bitMap.inline.hpp"
54 #include "utilities/growableArray.hpp"
55 #include "utilities/powerOfTwo.hpp"
56
57 //----------------------------GraphKit-----------------------------------------
58 // Main utility constructor.
59 GraphKit::GraphKit(JVMState* jvms)
60 : Phase(Phase::Parser),
61 _env(C->env()),
62 _gvn(*C->initial_gvn()),
63 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
64 {
65 _exceptions = jvms->map()->next_exception();
66 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
67 set_jvms(jvms);
68 }
69
70 // Private constructor for parser.
71 GraphKit::GraphKit()
72 : Phase(Phase::Parser),
73 _env(C->env()),
74 _gvn(*C->initial_gvn()),
75 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
76 {
77 _exceptions = nullptr;
78 set_map(nullptr);
79 DEBUG_ONLY(_sp = -99);
80 DEBUG_ONLY(set_bci(-99));
81 }
82
83
84
85 //---------------------------clean_stack---------------------------------------
86 // Clear away rubbish from the stack area of the JVM state.
87 // This destroys any arguments that may be waiting on the stack.
88 void GraphKit::clean_stack(int from_sp) {
89 SafePointNode* map = this->map();
90 JVMState* jvms = this->jvms();
91 int stk_size = jvms->stk_size();
92 int stkoff = jvms->stkoff();
93 Node* top = this->top();
94 for (int i = from_sp; i < stk_size; i++) {
95 if (map->in(stkoff + i) != top) {
96 map->set_req(stkoff + i, top);
97 }
98 }
99 }
100
101
102 //--------------------------------sync_jvms-----------------------------------
103 // Make sure our current jvms agrees with our parse state.
902 if (PrintMiscellaneous && (Verbose || WizardMode)) {
903 tty->print_cr("Zombie local %d: ", local);
904 jvms->dump();
905 }
906 return false;
907 }
908 }
909 }
910 return true;
911 }
912
913 #endif //ASSERT
914
915 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
916 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
917 ciMethod* cur_method = jvms->method();
918 int cur_bci = jvms->bci();
919 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
920 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
921 return Interpreter::bytecode_should_reexecute(code) ||
922 (is_anewarray && code == Bytecodes::_multianewarray);
923 // Reexecute _multianewarray bytecode which was replaced with
924 // sequence of [a]newarray. See Parse::do_multianewarray().
925 //
926 // Note: interpreter should not have it set since this optimization
927 // is limited by dimensions and guarded by flag so in some cases
928 // multianewarray() runtime calls will be generated and
929 // the bytecode should not be reexecutes (stack will not be reset).
930 } else {
931 return false;
932 }
933 }
934
935 // Helper function for adding JVMState and debug information to node
936 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
937 // Add the safepoint edges to the call (or other safepoint).
938
939 // Make sure dead locals are set to top. This
940 // should help register allocation time and cut down on the size
941 // of the deoptimization information.
942 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1018 uint p = debug_start; // walks forward in [debug_start, debug_end)
1019 uint j, k, l;
1020 SafePointNode* in_map = in_jvms->map();
1021 out_jvms->set_map(call);
1022
1023 if (can_prune_locals) {
1024 assert(in_jvms->method() == out_jvms->method(), "sanity");
1025 // If the current throw can reach an exception handler in this JVMS,
1026 // then we must keep everything live that can reach that handler.
1027 // As a quick and dirty approximation, we look for any handlers at all.
1028 if (in_jvms->method()->has_exception_handlers()) {
1029 can_prune_locals = false;
1030 }
1031 }
1032
1033 // Add the Locals
1034 k = in_jvms->locoff();
1035 l = in_jvms->loc_size();
1036 out_jvms->set_locoff(p);
1037 if (!can_prune_locals) {
1038 for (j = 0; j < l; j++)
1039 call->set_req(p++, in_map->in(k+j));
1040 } else {
1041 p += l; // already set to top above by add_req_batch
1042 }
1043
1044 // Add the Expression Stack
1045 k = in_jvms->stkoff();
1046 l = in_jvms->sp();
1047 out_jvms->set_stkoff(p);
1048 if (!can_prune_locals) {
1049 for (j = 0; j < l; j++)
1050 call->set_req(p++, in_map->in(k+j));
1051 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1052 // Divide stack into {S0,...,S1}, where S0 is set to top.
1053 uint s1 = stack_slots_not_pruned;
1054 stack_slots_not_pruned = 0; // for next iteration
1055 if (s1 > l) s1 = l;
1056 uint s0 = l - s1;
1057 p += s0; // skip the tops preinstalled by add_req_batch
1058 for (j = s0; j < l; j++)
1059 call->set_req(p++, in_map->in(k+j));
1060 } else {
1061 p += l; // already set to top above by add_req_batch
1062 }
1063
1064 // Add the Monitors
1065 k = in_jvms->monoff();
1066 l = in_jvms->mon_size();
1067 out_jvms->set_monoff(p);
1068 for (j = 0; j < l; j++)
1069 call->set_req(p++, in_map->in(k+j));
1070
1258 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1259 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1260 return _gvn.transform( new AndLNode(conv, mask) );
1261 }
1262
1263 Node* GraphKit::ConvL2I(Node* offset) {
1264 // short-circuit a common case
1265 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1266 if (offset_con != (jlong)Type::OffsetBot) {
1267 return intcon((int) offset_con);
1268 }
1269 return _gvn.transform( new ConvL2INode(offset));
1270 }
1271
1272 //-------------------------load_object_klass-----------------------------------
1273 Node* GraphKit::load_object_klass(Node* obj) {
1274 // Special-case a fresh allocation to avoid building nodes:
1275 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1276 if (akls != nullptr) return akls;
1277 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1278 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1279 }
1280
1281 //-------------------------load_array_length-----------------------------------
1282 Node* GraphKit::load_array_length(Node* array) {
1283 // Special-case a fresh allocation to avoid building nodes:
1284 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1285 Node *alen;
1286 if (alloc == nullptr) {
1287 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1288 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1289 } else {
1290 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1291 }
1292 return alen;
1293 }
1294
1295 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1296 const TypeOopPtr* oop_type,
1297 bool replace_length_in_map) {
1298 Node* length = alloc->Ideal_length();
1307 replace_in_map(length, ccast);
1308 }
1309 return ccast;
1310 }
1311 }
1312 return length;
1313 }
1314
1315 //------------------------------do_null_check----------------------------------
1316 // Helper function to do a null pointer check. Returned value is
1317 // the incoming address with null casted away. You are allowed to use the
1318 // not-null value only if you are control dependent on the test.
1319 #ifndef PRODUCT
1320 extern uint explicit_null_checks_inserted,
1321 explicit_null_checks_elided;
1322 #endif
1323 Node* GraphKit::null_check_common(Node* value, BasicType type,
1324 // optional arguments for variations:
1325 bool assert_null,
1326 Node* *null_control,
1327 bool speculative) {
1328 assert(!assert_null || null_control == nullptr, "not both at once");
1329 if (stopped()) return top();
1330 NOT_PRODUCT(explicit_null_checks_inserted++);
1331
1332 // Construct null check
1333 Node *chk = nullptr;
1334 switch(type) {
1335 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1336 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1337 case T_ARRAY : // fall through
1338 type = T_OBJECT; // simplify further tests
1339 case T_OBJECT : {
1340 const Type *t = _gvn.type( value );
1341
1342 const TypeOopPtr* tp = t->isa_oopptr();
1343 if (tp != nullptr && !tp->is_loaded()
1344 // Only for do_null_check, not any of its siblings:
1345 && !assert_null && null_control == nullptr) {
1346 // Usually, any field access or invocation on an unloaded oop type
1347 // will simply fail to link, since the statically linked class is
1348 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1349 // the static class is loaded but the sharper oop type is not.
1350 // Rather than checking for this obscure case in lots of places,
1351 // we simply observe that a null check on an unloaded class
1415 }
1416 Node *oldcontrol = control();
1417 set_control(cfg);
1418 Node *res = cast_not_null(value);
1419 set_control(oldcontrol);
1420 NOT_PRODUCT(explicit_null_checks_elided++);
1421 return res;
1422 }
1423 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1424 if (cfg == nullptr) break; // Quit at region nodes
1425 depth++;
1426 }
1427 }
1428
1429 //-----------
1430 // Branch to failure if null
1431 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1432 Deoptimization::DeoptReason reason;
1433 if (assert_null) {
1434 reason = Deoptimization::reason_null_assert(speculative);
1435 } else if (type == T_OBJECT) {
1436 reason = Deoptimization::reason_null_check(speculative);
1437 } else {
1438 reason = Deoptimization::Reason_div0_check;
1439 }
1440 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1441 // ciMethodData::has_trap_at will return a conservative -1 if any
1442 // must-be-null assertion has failed. This could cause performance
1443 // problems for a method after its first do_null_assert failure.
1444 // Consider using 'Reason_class_check' instead?
1445
1446 // To cause an implicit null check, we set the not-null probability
1447 // to the maximum (PROB_MAX). For an explicit check the probability
1448 // is set to a smaller value.
1449 if (null_control != nullptr || too_many_traps(reason)) {
1450 // probability is less likely
1451 ok_prob = PROB_LIKELY_MAG(3);
1452 } else if (!assert_null &&
1453 (ImplicitNullCheckThreshold > 0) &&
1454 method() != nullptr &&
1455 (method()->method_data()->trap_count(reason)
1489 }
1490
1491 if (assert_null) {
1492 // Cast obj to null on this path.
1493 replace_in_map(value, zerocon(type));
1494 return zerocon(type);
1495 }
1496
1497 // Cast obj to not-null on this path, if there is no null_control.
1498 // (If there is a null_control, a non-null value may come back to haunt us.)
1499 if (type == T_OBJECT) {
1500 Node* cast = cast_not_null(value, false);
1501 if (null_control == nullptr || (*null_control) == top())
1502 replace_in_map(value, cast);
1503 value = cast;
1504 }
1505
1506 return value;
1507 }
1508
1509
1510 //------------------------------cast_not_null----------------------------------
1511 // Cast obj to not-null on this path
1512 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1513 const Type *t = _gvn.type(obj);
1514 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1515 // Object is already not-null?
1516 if( t == t_not_null ) return obj;
1517
1518 Node* cast = new CastPPNode(control(), obj,t_not_null);
1519 cast = _gvn.transform( cast );
1520
1521 // Scan for instances of 'obj' in the current JVM mapping.
1522 // These instances are known to be not-null after the test.
1523 if (do_replace_in_map)
1524 replace_in_map(obj, cast);
1525
1526 return cast; // Return casted value
1527 }
1528
1529 // Sometimes in intrinsics, we implicitly know an object is not null
1530 // (there's no actual null check) so we can cast it to not null. In
1531 // the course of optimizations, the input to the cast can become null.
1532 // In that case that data path will die and we need the control path
1533 // to become dead as well to keep the graph consistent. So we have to
1534 // add a check for null for which one branch can't be taken. It uses
1535 // an OpaqueConstantBool node that will cause the check to be removed after loop
1536 // opts so the test goes away and the compiled code doesn't execute a
1537 // useless check.
1538 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1539 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1540 return value;
1541 }
1542 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1543 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1544 Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1545 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1546 _gvn.set_type(iff, iff->Value(&_gvn));
1585 //=============================================================================
1586 //--------------------------------memory---------------------------------------
1587 Node* GraphKit::memory(uint alias_idx) {
1588 MergeMemNode* mem = merged_memory();
1589 Node* p = mem->memory_at(alias_idx);
1590 assert(p != mem->empty_memory(), "empty");
1591 _gvn.set_type(p, Type::MEMORY); // must be mapped
1592 return p;
1593 }
1594
1595 //-----------------------------reset_memory------------------------------------
1596 Node* GraphKit::reset_memory() {
1597 Node* mem = map()->memory();
1598 // do not use this node for any more parsing!
1599 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1600 return _gvn.transform( mem );
1601 }
1602
1603 //------------------------------set_all_memory---------------------------------
1604 void GraphKit::set_all_memory(Node* newmem) {
1605 Node* mergemem = MergeMemNode::make(newmem);
1606 gvn().set_type_bottom(mergemem);
1607 map()->set_memory(mergemem);
1608 }
1609
1610 //------------------------------set_all_memory_call----------------------------
1611 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1612 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1613 set_all_memory(newmem);
1614 }
1615
1616 //=============================================================================
1617 //
1618 // parser factory methods for MemNodes
1619 //
1620 // These are layered on top of the factory methods in LoadNode and StoreNode,
1621 // and integrate with the parser's memory state and _gvn engine.
1622 //
1623
1624 // factory methods in "int adr_idx"
1625 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1626 MemNode::MemOrd mo,
1627 LoadNode::ControlDependency control_dependency,
1628 bool require_atomic_access,
1629 bool unaligned,
1630 bool mismatched,
1631 bool unsafe,
1632 uint8_t barrier_data) {
1633 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1634 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1635 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1636 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1637 Node* mem = memory(adr_idx);
1638 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1639 ld = _gvn.transform(ld);
1640 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1641 // Improve graph before escape analysis and boxing elimination.
1642 record_for_igvn(ld);
1643 if (ld->is_DecodeN()) {
1644 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1645 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1646 // a Phi). Recording such cases is still perfectly sound, but may be
1647 // unnecessary and result in some minor IGVN overhead.
1648 record_for_igvn(ld->in(1));
1649 }
1650 }
1651 return ld;
1652 }
1653
1654 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1655 MemNode::MemOrd mo,
1656 bool require_atomic_access,
1657 bool unaligned,
1658 bool mismatched,
1659 bool unsafe,
1673 if (unsafe) {
1674 st->as_Store()->set_unsafe_access();
1675 }
1676 st->as_Store()->set_barrier_data(barrier_data);
1677 st = _gvn.transform(st);
1678 set_memory(st, adr_idx);
1679 // Back-to-back stores can only remove intermediate store with DU info
1680 // so push on worklist for optimizer.
1681 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1682 record_for_igvn(st);
1683
1684 return st;
1685 }
1686
1687 Node* GraphKit::access_store_at(Node* obj,
1688 Node* adr,
1689 const TypePtr* adr_type,
1690 Node* val,
1691 const Type* val_type,
1692 BasicType bt,
1693 DecoratorSet decorators) {
1694 // Transformation of a value which could be null pointer (CastPP #null)
1695 // could be delayed during Parse (for example, in adjust_map_after_if()).
1696 // Execute transformation here to avoid barrier generation in such case.
1697 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1698 val = _gvn.makecon(TypePtr::NULL_PTR);
1699 }
1700
1701 if (stopped()) {
1702 return top(); // Dead path ?
1703 }
1704
1705 assert(val != nullptr, "not dead path");
1706
1707 C2AccessValuePtr addr(adr, adr_type);
1708 C2AccessValue value(val, val_type);
1709 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1710 if (access.is_raw()) {
1711 return _barrier_set->BarrierSetC2::store_at(access, value);
1712 } else {
1713 return _barrier_set->store_at(access, value);
1714 }
1715 }
1716
1717 Node* GraphKit::access_load_at(Node* obj, // containing obj
1718 Node* adr, // actual address to store val at
1719 const TypePtr* adr_type,
1720 const Type* val_type,
1721 BasicType bt,
1722 DecoratorSet decorators) {
1723 if (stopped()) {
1724 return top(); // Dead path ?
1725 }
1726
1727 SavedState old_state(this);
1728 C2AccessValuePtr addr(adr, adr_type);
1729 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1730 Node* load;
1731 if (access.is_raw()) {
1732 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1733 } else {
1734 load = _barrier_set->load_at(access, val_type);
1735 }
1736
1737 // Restore the previous state only if the load got folded to a constant
1738 // and we can discard any barriers that might have been added.
1739 if (load == nullptr || !load->is_Con()) {
1740 old_state.discard();
1741 }
1742 return load;
1743 }
1744
1745 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1746 const Type* val_type,
1747 BasicType bt,
1748 DecoratorSet decorators) {
1749 if (stopped()) {
1831 Node* new_val,
1832 const Type* value_type,
1833 BasicType bt,
1834 DecoratorSet decorators) {
1835 C2AccessValuePtr addr(adr, adr_type);
1836 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1837 if (access.is_raw()) {
1838 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1839 } else {
1840 return _barrier_set->atomic_add_at(access, new_val, value_type);
1841 }
1842 }
1843
1844 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1845 return _barrier_set->clone(this, src, dst, size, is_array);
1846 }
1847
1848 //-------------------------array_element_address-------------------------
1849 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1850 const TypeInt* sizetype, Node* ctrl) {
1851 uint shift = exact_log2(type2aelembytes(elembt));
1852 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1853
1854 // short-circuit a common case (saves lots of confusing waste motion)
1855 jint idx_con = find_int_con(idx, -1);
1856 if (idx_con >= 0) {
1857 intptr_t offset = header + ((intptr_t)idx_con << shift);
1858 return basic_plus_adr(ary, offset);
1859 }
1860
1861 // must be correct type for alignment purposes
1862 Node* base = basic_plus_adr(ary, header);
1863 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1864 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1865 return basic_plus_adr(ary, base, scale);
1866 }
1867
1868 //-------------------------load_array_element-------------------------
1869 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1870 const Type* elemtype = arytype->elem();
1871 BasicType elembt = elemtype->array_element_basic_type();
1872 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1873 if (elembt == T_NARROWOOP) {
1874 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1875 }
1876 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1877 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1878 return ld;
1879 }
1880
1881 //-------------------------set_arguments_for_java_call-------------------------
1882 // Arguments (pre-popped from the stack) are taken from the JVMS.
1883 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1884 // Add the call arguments:
1885 uint nargs = call->method()->arg_size();
1886 for (uint i = 0; i < nargs; i++) {
1887 Node* arg = argument(i);
1888 call->init_req(i + TypeFunc::Parms, arg);
1889 }
1890 }
1891
1892 //---------------------------set_edges_for_java_call---------------------------
1893 // Connect a newly created call into the current JVMS.
1894 // A return value node (if any) is returned from set_edges_for_java_call.
1895 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1896
1897 // Add the predefined inputs:
1898 call->init_req( TypeFunc::Control, control() );
1899 call->init_req( TypeFunc::I_O , i_o() );
1900 call->init_req( TypeFunc::Memory , reset_memory() );
1901 call->init_req( TypeFunc::FramePtr, frameptr() );
1902 call->init_req( TypeFunc::ReturnAdr, top() );
1903
1904 add_safepoint_edges(call, must_throw);
1905
1906 Node* xcall = _gvn.transform(call);
1907
1908 if (xcall == top()) {
1909 set_control(top());
1910 return;
1911 }
1912 assert(xcall == call, "call identity is stable");
1913
1914 // Re-use the current map to produce the result.
1915
1916 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1917 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1918 set_all_memory_call(xcall, separate_io_proj);
1919
1920 //return xcall; // no need, caller already has it
1921 }
1922
1923 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1924 if (stopped()) return top(); // maybe the call folded up?
1925
1926 // Capture the return value, if any.
1927 Node* ret;
1928 if (call->method() == nullptr ||
1929 call->method()->return_type()->basic_type() == T_VOID)
1930 ret = top();
1931 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1932
1933 // Note: Since any out-of-line call can produce an exception,
1934 // we always insert an I_O projection from the call into the result.
1935
1936 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1937
1938 if (separate_io_proj) {
1939 // The caller requested separate projections be used by the fall
1940 // through and exceptional paths, so replace the projections for
1941 // the fall through path.
1942 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1943 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1944 }
1945 return ret;
1946 }
1947
1948 //--------------------set_predefined_input_for_runtime_call--------------------
1949 // Reading and setting the memory state is way conservative here.
1950 // The real problem is that I am not doing real Type analysis on memory,
1951 // so I cannot distinguish card mark stores from other stores. Across a GC
1952 // point the Store Barrier and the card mark memory has to agree. I cannot
1953 // have a card mark store and its barrier split across the GC point from
1954 // either above or below. Here I get that to happen by reading ALL of memory.
1955 // A better answer would be to separate out card marks from other memory.
1956 // For now, return the input memory state, so that it can be reused
1957 // after the call, if this call has restricted memory effects.
1958 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1959 // Set fixed predefined input arguments
1960 call->init_req(TypeFunc::Control, control());
1961 call->init_req(TypeFunc::I_O, top()); // does no i/o
1962 call->init_req(TypeFunc::ReturnAdr, top());
1963 if (call->is_CallLeafPure()) {
1964 call->init_req(TypeFunc::Memory, top());
2026 if (use->is_MergeMem()) {
2027 wl.push(use);
2028 }
2029 }
2030 }
2031
2032 // Replace the call with the current state of the kit.
2033 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2034 JVMState* ejvms = nullptr;
2035 if (has_exceptions()) {
2036 ejvms = transfer_exceptions_into_jvms();
2037 }
2038
2039 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2040 ReplacedNodes replaced_nodes_exception;
2041 Node* ex_ctl = top();
2042
2043 SafePointNode* final_state = stop();
2044
2045 // Find all the needed outputs of this call
2046 CallProjections callprojs;
2047 call->extract_projections(&callprojs, true, do_asserts);
2048
2049 Unique_Node_List wl;
2050 Node* init_mem = call->in(TypeFunc::Memory);
2051 Node* final_mem = final_state->in(TypeFunc::Memory);
2052 Node* final_ctl = final_state->in(TypeFunc::Control);
2053 Node* final_io = final_state->in(TypeFunc::I_O);
2054
2055 // Replace all the old call edges with the edges from the inlining result
2056 if (callprojs.fallthrough_catchproj != nullptr) {
2057 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
2058 }
2059 if (callprojs.fallthrough_memproj != nullptr) {
2060 if (final_mem->is_MergeMem()) {
2061 // Parser's exits MergeMem was not transformed but may be optimized
2062 final_mem = _gvn.transform(final_mem);
2063 }
2064 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
2065 add_mergemem_users_to_worklist(wl, final_mem);
2066 }
2067 if (callprojs.fallthrough_ioproj != nullptr) {
2068 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
2069 }
2070
2071 // Replace the result with the new result if it exists and is used
2072 if (callprojs.resproj != nullptr && result != nullptr) {
2073 C->gvn_replace_by(callprojs.resproj, result);
2074 }
2075
2076 if (ejvms == nullptr) {
2077 // No exception edges to simply kill off those paths
2078 if (callprojs.catchall_catchproj != nullptr) {
2079 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
2080 }
2081 if (callprojs.catchall_memproj != nullptr) {
2082 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
2083 }
2084 if (callprojs.catchall_ioproj != nullptr) {
2085 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
2086 }
2087 // Replace the old exception object with top
2088 if (callprojs.exobj != nullptr) {
2089 C->gvn_replace_by(callprojs.exobj, C->top());
2090 }
2091 } else {
2092 GraphKit ekit(ejvms);
2093
2094 // Load my combined exception state into the kit, with all phis transformed:
2095 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2096 replaced_nodes_exception = ex_map->replaced_nodes();
2097
2098 Node* ex_oop = ekit.use_exception_state(ex_map);
2099
2100 if (callprojs.catchall_catchproj != nullptr) {
2101 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2102 ex_ctl = ekit.control();
2103 }
2104 if (callprojs.catchall_memproj != nullptr) {
2105 Node* ex_mem = ekit.reset_memory();
2106 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2107 add_mergemem_users_to_worklist(wl, ex_mem);
2108 }
2109 if (callprojs.catchall_ioproj != nullptr) {
2110 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2111 }
2112
2113 // Replace the old exception object with the newly created one
2114 if (callprojs.exobj != nullptr) {
2115 C->gvn_replace_by(callprojs.exobj, ex_oop);
2116 }
2117 }
2118
2119 // Disconnect the call from the graph
2120 call->disconnect_inputs(C);
2121 C->gvn_replace_by(call, C->top());
2122
2123 // Clean up any MergeMems that feed other MergeMems since the
2124 // optimizer doesn't like that.
2125 while (wl.size() > 0) {
2126 _gvn.transform(wl.pop());
2127 }
2128
2129 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2130 replaced_nodes.apply(C, final_ctl);
2131 }
2132 if (!ex_ctl->is_top() && do_replaced_nodes) {
2133 replaced_nodes_exception.apply(C, ex_ctl);
2134 }
2135 }
2136
2137
2138 //------------------------------increment_counter------------------------------
2139 // for statistics: increment a VM counter by 1
2140
2141 void GraphKit::increment_counter(address counter_addr) {
2142 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2143 increment_counter(adr1);
2144 }
2145
2146 void GraphKit::increment_counter(Node* counter_addr) {
2147 Node* ctrl = control();
2148 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2149 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2150 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2151 }
2152
2153 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2154 Node* halt = new HaltNode(ctrl, frameptr, reason
2155 PRODUCT_ONLY(COMMA generate_code_in_product));
2156 halt = _gvn.transform(halt);
2157 root()->add_req(halt);
2158 }
2159
2160 //------------------------------uncommon_trap----------------------------------
2161 // Bail out to the interpreter in mid-method. Implemented by calling the
2162 // uncommon_trap blob. This helper function inserts a runtime call with the
2163 // right debug info.
2164 Node* GraphKit::uncommon_trap(int trap_request,
2165 ciKlass* klass, const char* comment,
2166 bool must_throw,
2167 bool keep_exact_action) {
2168 if (failing_internal()) {
2169 stop();
2170 }
2171 if (stopped()) return nullptr; // trap reachable?
2172
2173 // Note: If ProfileTraps is true, and if a deopt. actually
2174 // occurs here, the runtime will make sure an MDO exists. There is
2175 // no need to call method()->ensure_method_data() at this point.
2176
2177 // Set the stack pointer to the right value for reexecution:
2315
2316 /**
2317 * Record profiling data exact_kls for Node n with the type system so
2318 * that it can propagate it (speculation)
2319 *
2320 * @param n node that the type applies to
2321 * @param exact_kls type from profiling
2322 * @param maybe_null did profiling see null?
2323 *
2324 * @return node with improved type
2325 */
2326 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2327 const Type* current_type = _gvn.type(n);
2328 assert(UseTypeSpeculation, "type speculation must be on");
2329
2330 const TypePtr* speculative = current_type->speculative();
2331
2332 // Should the klass from the profile be recorded in the speculative type?
2333 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2334 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2335 const TypeOopPtr* xtype = tklass->as_instance_type();
2336 assert(xtype->klass_is_exact(), "Should be exact");
2337 // Any reason to believe n is not null (from this profiling or a previous one)?
2338 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2339 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2340 // record the new speculative type's depth
2341 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2342 speculative = speculative->with_inline_depth(jvms()->depth());
2343 } else if (current_type->would_improve_ptr(ptr_kind)) {
2344 // Profiling report that null was never seen so we can change the
2345 // speculative type to non null ptr.
2346 if (ptr_kind == ProfileAlwaysNull) {
2347 speculative = TypePtr::NULL_PTR;
2348 } else {
2349 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2350 const TypePtr* ptr = TypePtr::NOTNULL;
2351 if (speculative != nullptr) {
2352 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2353 } else {
2354 speculative = ptr;
2355 }
2356 }
2357 }
2358
2359 if (speculative != current_type->speculative()) {
2360 // Build a type with a speculative type (what we think we know
2361 // about the type but will need a guard when we use it)
2362 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2363 // We're changing the type, we need a new CheckCast node to carry
2364 // the new type. The new type depends on the control: what
2365 // profiling tells us is only valid from here as far as we can
2366 // tell.
2367 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2368 cast = _gvn.transform(cast);
2369 replace_in_map(n, cast);
2370 n = cast;
2371 }
2372
2373 return n;
2374 }
2375
2376 /**
2377 * Record profiling data from receiver profiling at an invoke with the
2378 * type system so that it can propagate it (speculation)
2379 *
2380 * @param n receiver node
2381 *
2382 * @return node with improved type
2383 */
2384 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2385 if (!UseTypeSpeculation) {
2386 return n;
2387 }
2388 ciKlass* exact_kls = profile_has_unique_klass();
2389 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2390 if ((java_bc() == Bytecodes::_checkcast ||
2391 java_bc() == Bytecodes::_instanceof ||
2392 java_bc() == Bytecodes::_aastore) &&
2393 method()->method_data()->is_mature()) {
2394 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2395 if (data != nullptr) {
2396 if (!data->as_BitData()->null_seen()) {
2397 ptr_kind = ProfileNeverNull;
2398 } else {
2399 if (TypeProfileCasts) {
2400 assert(data->is_ReceiverTypeData(), "bad profile data type");
2401 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2402 uint i = 0;
2403 for (; i < call->row_limit(); i++) {
2404 ciKlass* receiver = call->receiver(i);
2405 if (receiver != nullptr) {
2406 break;
2407 }
2408 }
2409 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2410 }
2411 }
2412 }
2413 }
2414 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2415 }
2416
2417 /**
2418 * Record profiling data from argument profiling at an invoke with the
2419 * type system so that it can propagate it (speculation)
2420 *
2421 * @param dest_method target method for the call
2422 * @param bc what invoke bytecode is this?
2423 */
2424 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2425 if (!UseTypeSpeculation) {
2426 return;
2427 }
2428 const TypeFunc* tf = TypeFunc::make(dest_method);
2429 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2430 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2431 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2432 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2433 if (is_reference_type(targ->basic_type())) {
2434 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2435 ciKlass* better_type = nullptr;
2436 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2437 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2438 }
2439 i++;
2440 }
2441 }
2442 }
2443
2444 /**
2445 * Record profiling data from parameter profiling at an invoke with
2446 * the type system so that it can propagate it (speculation)
2447 */
2448 void GraphKit::record_profiled_parameters_for_speculation() {
2449 if (!UseTypeSpeculation) {
2450 return;
2451 }
2452 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2572 // The first null ends the list.
2573 Node* parm0, Node* parm1,
2574 Node* parm2, Node* parm3,
2575 Node* parm4, Node* parm5,
2576 Node* parm6, Node* parm7) {
2577 assert(call_addr != nullptr, "must not call null targets");
2578
2579 // Slow-path call
2580 bool is_leaf = !(flags & RC_NO_LEAF);
2581 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2582 if (call_name == nullptr) {
2583 assert(!is_leaf, "must supply name for leaf");
2584 call_name = OptoRuntime::stub_name(call_addr);
2585 }
2586 CallNode* call;
2587 if (!is_leaf) {
2588 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2589 } else if (flags & RC_NO_FP) {
2590 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2591 } else if (flags & RC_VECTOR){
2592 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2593 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2594 } else if (flags & RC_PURE) {
2595 assert(adr_type == nullptr, "pure call does not touch memory");
2596 call = new CallLeafPureNode(call_type, call_addr, call_name);
2597 } else {
2598 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2599 }
2600
2601 // The following is similar to set_edges_for_java_call,
2602 // except that the memory effects of the call are restricted to AliasIdxRaw.
2603
2604 // Slow path call has no side-effects, uses few values
2605 bool wide_in = !(flags & RC_NARROW_MEM);
2606 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2607
2608 Node* prev_mem = nullptr;
2609 if (wide_in) {
2610 prev_mem = set_predefined_input_for_runtime_call(call);
2611 } else {
2612 assert(!wide_out, "narrow in => narrow out");
2613 Node* narrow_mem = memory(adr_type);
2614 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2615 }
2616
2617 // Hook each parm in order. Stop looking at the first null.
2618 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2619 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2620 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2621 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2622 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2623 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2624 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2625 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2626 /* close each nested if ===> */ } } } } } } } }
2627 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2628
2629 if (!is_leaf) {
2630 // Non-leaves can block and take safepoints:
2631 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2632 }
2633 // Non-leaves can throw exceptions:
2634 if (has_io) {
2635 call->set_req(TypeFunc::I_O, i_o());
2636 }
2637
2638 if (flags & RC_UNCOMMON) {
2639 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2640 // (An "if" probability corresponds roughly to an unconditional count.
2641 // Sort of.)
2642 call->set_cnt(PROB_UNLIKELY_MAG(4));
2643 }
2644
2645 Node* c = _gvn.transform(call);
2646 assert(c == call, "cannot disappear");
2647
2655
2656 if (has_io) {
2657 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2658 }
2659 return call;
2660
2661 }
2662
2663 // i2b
2664 Node* GraphKit::sign_extend_byte(Node* in) {
2665 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2666 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2667 }
2668
2669 // i2s
2670 Node* GraphKit::sign_extend_short(Node* in) {
2671 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2672 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2673 }
2674
2675 //------------------------------merge_memory-----------------------------------
2676 // Merge memory from one path into the current memory state.
2677 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2678 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2679 Node* old_slice = mms.force_memory();
2680 Node* new_slice = mms.memory2();
2681 if (old_slice != new_slice) {
2682 PhiNode* phi;
2683 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2684 if (mms.is_empty()) {
2685 // clone base memory Phi's inputs for this memory slice
2686 assert(old_slice == mms.base_memory(), "sanity");
2687 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2688 _gvn.set_type(phi, Type::MEMORY);
2689 for (uint i = 1; i < phi->req(); i++) {
2690 phi->init_req(i, old_slice->in(i));
2691 }
2692 } else {
2693 phi = old_slice->as_Phi(); // Phi was generated already
2694 }
2751 gvn.transform(iff);
2752 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2753 return iff;
2754 }
2755
2756 //-------------------------------gen_subtype_check-----------------------------
2757 // Generate a subtyping check. Takes as input the subtype and supertype.
2758 // Returns 2 values: sets the default control() to the true path and returns
2759 // the false path. Only reads invariant memory; sets no (visible) memory.
2760 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2761 // but that's not exposed to the optimizer. This call also doesn't take in an
2762 // Object; if you wish to check an Object you need to load the Object's class
2763 // prior to coming here.
2764 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2765 ciMethod* method, int bci) {
2766 Compile* C = gvn.C;
2767 if ((*ctrl)->is_top()) {
2768 return C->top();
2769 }
2770
2771 // Fast check for identical types, perhaps identical constants.
2772 // The types can even be identical non-constants, in cases
2773 // involving Array.newInstance, Object.clone, etc.
2774 if (subklass == superklass)
2775 return C->top(); // false path is dead; no test needed.
2776
2777 if (gvn.type(superklass)->singleton()) {
2778 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2779 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2780
2781 // In the common case of an exact superklass, try to fold up the
2782 // test before generating code. You may ask, why not just generate
2783 // the code and then let it fold up? The answer is that the generated
2784 // code will necessarily include null checks, which do not always
2785 // completely fold away. If they are also needless, then they turn
2786 // into a performance loss. Example:
2787 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2788 // Here, the type of 'fa' is often exact, so the store check
2789 // of fa[1]=x will fold up, without testing the nullness of x.
2790 //
2791 // At macro expansion, we would have already folded the SubTypeCheckNode
2792 // being expanded here because we always perform the static sub type
2793 // check in SubTypeCheckNode::sub() regardless of whether
2794 // StressReflectiveCode is set or not. We can therefore skip this
2795 // static check when StressReflectiveCode is on.
2796 switch (C->static_subtype_check(superk, subk)) {
2797 case Compile::SSC_always_false:
2798 {
2799 Node* always_fail = *ctrl;
2800 *ctrl = gvn.C->top();
2801 return always_fail;
2802 }
2803 case Compile::SSC_always_true:
2804 return C->top();
2805 case Compile::SSC_easy_test:
2806 {
2807 // Just do a direct pointer compare and be done.
2808 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2809 *ctrl = gvn.transform(new IfTrueNode(iff));
2810 return gvn.transform(new IfFalseNode(iff));
2811 }
2812 case Compile::SSC_full_test:
2813 break;
2814 default:
2815 ShouldNotReachHere();
2816 }
2817 }
2818
2819 // %%% Possible further optimization: Even if the superklass is not exact,
2820 // if the subklass is the unique subtype of the superklass, the check
2821 // will always succeed. We could leave a dependency behind to ensure this.
2822
2823 // First load the super-klass's check-offset
2824 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2825 Node* m = C->immutable_memory();
2826 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2827 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2828 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2866 gvn.record_for_igvn(r_ok_subtype);
2867
2868 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2869 // SubTypeCheck node
2870 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2871 ciCallProfile profile = method->call_profile_at_bci(bci);
2872 float total_prob = 0;
2873 for (int i = 0; profile.has_receiver(i); ++i) {
2874 float prob = profile.receiver_prob(i);
2875 total_prob += prob;
2876 }
2877 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2878 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2879 for (int i = 0; profile.has_receiver(i); ++i) {
2880 ciKlass* klass = profile.receiver(i);
2881 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2882 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2883 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2884 continue;
2885 }
2886 float prob = profile.receiver_prob(i);
2887 ConNode* klass_node = gvn.makecon(klass_t);
2888 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2889 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2890
2891 if (result == Compile::SSC_always_true) {
2892 r_ok_subtype->add_req(iftrue);
2893 } else {
2894 assert(result == Compile::SSC_always_false, "");
2895 r_not_subtype->add_req(iftrue);
2896 }
2897 *ctrl = gvn.transform(new IfFalseNode(iff));
2898 }
2899 }
2900 }
2901
2902 // See if we get an immediate positive hit. Happens roughly 83% of the
2903 // time. Test to see if the value loaded just previously from the subklass
2904 // is exactly the superklass.
2905 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2919 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
2920 }
2921 return not_subtype_ctrl;
2922 }
2923
2924 r_ok_subtype->init_req(1, iftrue1);
2925
2926 // Check for immediate negative hit. Happens roughly 11% of the time (which
2927 // is roughly 63% of the remaining cases). Test to see if the loaded
2928 // check-offset points into the subklass display list or the 1-element
2929 // cache. If it points to the display (and NOT the cache) and the display
2930 // missed then it's not a subtype.
2931 Node *cacheoff = gvn.intcon(cacheoff_con);
2932 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2933 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2934 *ctrl = gvn.transform(new IfFalseNode(iff2));
2935
2936 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2937 // No performance impact (too rare) but allows sharing of secondary arrays
2938 // which has some footprint reduction.
2939 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2940 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2941 *ctrl = gvn.transform(new IfFalseNode(iff3));
2942
2943 // -- Roads not taken here: --
2944 // We could also have chosen to perform the self-check at the beginning
2945 // of this code sequence, as the assembler does. This would not pay off
2946 // the same way, since the optimizer, unlike the assembler, can perform
2947 // static type analysis to fold away many successful self-checks.
2948 // Non-foldable self checks work better here in second position, because
2949 // the initial primary superclass check subsumes a self-check for most
2950 // types. An exception would be a secondary type like array-of-interface,
2951 // which does not appear in its own primary supertype display.
2952 // Finally, we could have chosen to move the self-check into the
2953 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2954 // dependent manner. But it is worthwhile to have the check here,
2955 // where it can be perhaps be optimized. The cost in code space is
2956 // small (register compare, branch).
2957
2958 // Now do a linear scan of the secondary super-klass array. Again, no real
2959 // performance impact (too rare) but it's gotta be done.
2960 // Since the code is rarely used, there is no penalty for moving it
2961 // out of line, and it can only improve I-cache density.
2962 // The decision to inline or out-of-line this final check is platform
2963 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2964 Node* psc = gvn.transform(
2965 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2966
2967 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2968 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2969 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2970
2971 // Return false path; set default control to true path.
2972 *ctrl = gvn.transform(r_ok_subtype);
2973 return gvn.transform(r_not_subtype);
2974 }
2975
2976 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2977 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2978 if (expand_subtype_check) {
2979 MergeMemNode* mem = merged_memory();
2980 Node* ctrl = control();
2981 Node* subklass = obj_or_subklass;
2982 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2983 subklass = load_object_klass(obj_or_subklass);
2984 }
2985
2986 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2987 set_control(ctrl);
2988 return n;
2989 }
2990
2991 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2992 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2993 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2994 set_control(_gvn.transform(new IfTrueNode(iff)));
2995 return _gvn.transform(new IfFalseNode(iff));
2996 }
2997
2998 // Profile-driven exact type check:
2999 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3000 float prob,
3001 Node* *casted_receiver) {
3002 assert(!klass->is_interface(), "no exact type check on interfaces");
3003
3004 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3005 Node* recv_klass = load_object_klass(receiver);
3006 Node* want_klass = makecon(tklass);
3007 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3008 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3009 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3010 set_control( _gvn.transform(new IfTrueNode (iff)));
3011 Node* fail = _gvn.transform(new IfFalseNode(iff));
3012
3013 if (!stopped()) {
3014 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3015 const TypeOopPtr* recvx_type = tklass->as_instance_type();
3016 assert(recvx_type->klass_is_exact(), "");
3017
3018 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
3019 // Subsume downstream occurrences of receiver with a cast to
3020 // recv_xtype, since now we know what the type will be.
3021 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
3022 (*casted_receiver) = _gvn.transform(cast);
3023 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3024 // (User must make the replace_in_map call.)
3025 }
3026 }
3027
3028 return fail;
3029 }
3030
3031 //------------------------------subtype_check_receiver-------------------------
3032 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3033 Node** casted_receiver) {
3034 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3035 Node* want_klass = makecon(tklass);
3036
3037 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3038
3039 // Ignore interface type information until interface types are properly tracked.
3040 if (!stopped() && !klass->is_interface()) {
3041 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3042 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3043 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3044 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3045 (*casted_receiver) = _gvn.transform(cast);
3046 }
3047 }
3048
3049 return slow_ctl;
3050 }
3051
3052 //------------------------------seems_never_null-------------------------------
3053 // Use null_seen information if it is available from the profile.
3054 // If we see an unexpected null at a type check we record it and force a
3055 // recompile; the offending check will be recompiled to handle nulls.
3056 // If we see several offending BCIs, then all checks in the
3057 // method will be recompiled.
3058 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3059 speculating = !_gvn.type(obj)->speculative_maybe_null();
3060 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3061 if (UncommonNullCast // Cutout for this technique
3062 && obj != null() // And not the -Xcomp stupid case?
3063 && !too_many_traps(reason)
3064 ) {
3065 if (speculating) {
3134
3135 //------------------------maybe_cast_profiled_receiver-------------------------
3136 // If the profile has seen exactly one type, narrow to exactly that type.
3137 // Subsequent type checks will always fold up.
3138 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3139 const TypeKlassPtr* require_klass,
3140 ciKlass* spec_klass,
3141 bool safe_for_replace) {
3142 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3143
3144 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3145
3146 // Make sure we haven't already deoptimized from this tactic.
3147 if (too_many_traps_or_recompiles(reason))
3148 return nullptr;
3149
3150 // (No, this isn't a call, but it's enough like a virtual call
3151 // to use the same ciMethod accessor to get the profile info...)
3152 // If we have a speculative type use it instead of profiling (which
3153 // may not help us)
3154 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3155 if (exact_kls != nullptr) {// no cast failures here
3156 if (require_klass == nullptr ||
3157 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3158 // If we narrow the type to match what the type profile sees or
3159 // the speculative type, we can then remove the rest of the
3160 // cast.
3161 // This is a win, even if the exact_kls is very specific,
3162 // because downstream operations, such as method calls,
3163 // will often benefit from the sharper type.
3164 Node* exact_obj = not_null_obj; // will get updated in place...
3165 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3166 &exact_obj);
3167 { PreserveJVMState pjvms(this);
3168 set_control(slow_ctl);
3169 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3170 }
3171 if (safe_for_replace) {
3172 replace_in_map(not_null_obj, exact_obj);
3173 }
3174 return exact_obj;
3264
3265 // If not_null_obj is dead, only null-path is taken
3266 if (stopped()) { // Doing instance-of on a null?
3267 set_control(null_ctl);
3268 return intcon(0);
3269 }
3270 region->init_req(_null_path, null_ctl);
3271 phi ->init_req(_null_path, intcon(0)); // Set null path value
3272 if (null_ctl == top()) {
3273 // Do this eagerly, so that pattern matches like is_diamond_phi
3274 // will work even during parsing.
3275 assert(_null_path == PATH_LIMIT-1, "delete last");
3276 region->del_req(_null_path);
3277 phi ->del_req(_null_path);
3278 }
3279
3280 // Do we know the type check always succeed?
3281 bool known_statically = false;
3282 if (improved_klass_ptr_type->singleton()) {
3283 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3284 if (subk->is_loaded()) {
3285 int static_res = C->static_subtype_check(improved_klass_ptr_type, subk);
3286 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3287 }
3288 }
3289
3290 if (!known_statically) {
3291 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3292 // We may not have profiling here or it may not help us. If we
3293 // have a speculative type use it to perform an exact cast.
3294 ciKlass* spec_obj_type = obj_type->speculative_type();
3295 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3296 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3297 if (stopped()) { // Profile disagrees with this path.
3298 set_control(null_ctl); // Null is the only remaining possibility.
3299 return intcon(0);
3300 }
3301 if (cast_obj != nullptr) {
3302 not_null_obj = cast_obj;
3303 }
3304 }
3324 record_for_igvn(region);
3325
3326 // If we know the type check always succeeds then we don't use the
3327 // profiling data at this bytecode. Don't lose it, feed it to the
3328 // type system as a speculative type.
3329 if (safe_for_replace) {
3330 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3331 replace_in_map(obj, casted_obj);
3332 }
3333
3334 return _gvn.transform(phi);
3335 }
3336
3337 //-------------------------------gen_checkcast---------------------------------
3338 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3339 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3340 // uncommon-trap paths work. Adjust stack after this call.
3341 // If failure_control is supplied and not null, it is filled in with
3342 // the control edge for the cast failure. Otherwise, an appropriate
3343 // uncommon trap or exception is thrown.
3344 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3345 Node* *failure_control) {
3346 kill_dead_locals(); // Benefit all the uncommon traps
3347 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3348 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3349 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3350
3351 // Fast cutout: Check the case that the cast is vacuously true.
3352 // This detects the common cases where the test will short-circuit
3353 // away completely. We do this before we perform the null check,
3354 // because if the test is going to turn into zero code, we don't
3355 // want a residual null check left around. (Causes a slowdown,
3356 // for example, in some objArray manipulations, such as a[i]=a[j].)
3357 if (improved_klass_ptr_type->singleton()) {
3358 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3359 if (objtp != nullptr) {
3360 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3361 case Compile::SSC_always_true:
3362 // If we know the type check always succeed then we don't use
3363 // the profiling data at this bytecode. Don't lose it, feed it
3364 // to the type system as a speculative type.
3365 return record_profiled_receiver_for_speculation(obj);
3366 case Compile::SSC_always_false:
3367 // It needs a null check because a null will *pass* the cast check.
3368 // A non-null value will always produce an exception.
3369 if (!objtp->maybe_null()) {
3370 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3371 Deoptimization::DeoptReason reason = is_aastore ?
3372 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3373 builtin_throw(reason);
3374 return top();
3375 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3376 return null_assert(obj);
3377 }
3378 break; // Fall through to full check
3379 default:
3380 break;
3381 }
3382 }
3383 }
3384
3385 ciProfileData* data = nullptr;
3386 bool safe_for_replace = false;
3387 if (failure_control == nullptr) { // use MDO in regular case only
3388 assert(java_bc() == Bytecodes::_aastore ||
3389 java_bc() == Bytecodes::_checkcast,
3390 "interpreter profiles type checks only for these BCs");
3391 data = method()->method_data()->bci_to_data(bci());
3392 safe_for_replace = true;
3393 }
3394
3395 // Make the merge point
3396 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3397 RegionNode* region = new RegionNode(PATH_LIMIT);
3398 Node* phi = new PhiNode(region, toop);
3399 C->set_has_split_ifs(true); // Has chance for split-if optimization
3400
3401 // Use null-cast information if it is available
3402 bool speculative_not_null = false;
3403 bool never_see_null = ((failure_control == nullptr) // regular case only
3404 && seems_never_null(obj, data, speculative_not_null));
3405
3406 // Null check; get casted pointer; set region slot 3
3407 Node* null_ctl = top();
3408 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, false /*safe_for_replace*/, speculative_not_null);
3409
3410 // If not_null_obj is dead, only null-path is taken
3411 if (stopped()) { // Doing instance-of on a null?
3412 set_control(null_ctl);
3413 return null();
3414 }
3415 region->init_req(_null_path, null_ctl);
3416 phi ->init_req(_null_path, null()); // Set null path value
3417 if (null_ctl == top()) {
3418 // Do this eagerly, so that pattern matches like is_diamond_phi
3419 // will work even during parsing.
3420 assert(_null_path == PATH_LIMIT-1, "delete last");
3421 region->del_req(_null_path);
3422 phi ->del_req(_null_path);
3423 }
3424
3425 Node* cast_obj = nullptr;
3426 if (improved_klass_ptr_type->klass_is_exact()) {
3427 // The following optimization tries to statically cast the speculative type of the object
3428 // (for example obtained during profiling) to the type of the superklass and then do a
3429 // dynamic check that the type of the object is what we expect. To work correctly
3430 // for checkcast and aastore the type of superklass should be exact.
3431 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3432 // We may not have profiling here or it may not help us. If we have
3433 // a speculative type use it to perform an exact cast.
3434 ciKlass* spec_obj_type = obj_type->speculative_type();
3435 if (spec_obj_type != nullptr || data != nullptr) {
3436 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, false /*safe_for_replace*/);
3437 if (cast_obj != nullptr) {
3438 if (failure_control != nullptr) // failure is now impossible
3439 (*failure_control) = top();
3440 // adjust the type of the phi to the exact klass:
3441 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3442 }
3443 }
3444 }
3445
3446 if (cast_obj == nullptr) {
3447 // Generate the subtype check
3448 Node* improved_superklass = superklass;
3449 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3450 improved_superklass = makecon(improved_klass_ptr_type);
3451 }
3452 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3453
3454 // Plug in success path into the merge
3455 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3456 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3457 if (failure_control == nullptr) {
3458 if (not_subtype_ctrl != top()) { // If failure is possible
3459 PreserveJVMState pjvms(this);
3460 set_control(not_subtype_ctrl);
3461 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3462 Deoptimization::DeoptReason reason = is_aastore ?
3463 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3464 builtin_throw(reason);
3465 }
3466 } else {
3467 (*failure_control) = not_subtype_ctrl;
3468 }
3469 }
3470
3471 region->init_req(_obj_path, control());
3472 phi ->init_req(_obj_path, cast_obj);
3473
3474 // Return final merged results
3475 set_control( _gvn.transform(region) );
3476 record_for_igvn(region);
3477
3478 // A merge of null or Casted-NotNull obj
3479 Node* res = _gvn.transform(phi);
3480 res = record_profiled_receiver_for_speculation(res);
3481 if (safe_for_replace) {
3482 replace_in_map(obj, res);
3483 }
3484 return res;
3485 }
3486
3487 //------------------------------next_monitor-----------------------------------
3488 // What number should be given to the next monitor?
3489 int GraphKit::next_monitor() {
3490 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3491 int next = current + C->sync_stack_slots();
3492 // Keep the toplevel high water mark current:
3493 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3494 return current;
3495 }
3496
3497 //------------------------------insert_mem_bar---------------------------------
3498 // Memory barrier to avoid floating things around
3499 // The membar serves as a pinch point between both control and all memory slices.
3500 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3501 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3502 mb->init_req(TypeFunc::Control, control());
3503 mb->init_req(TypeFunc::Memory, reset_memory());
3504 Node* membar = _gvn.transform(mb);
3505 record_for_igvn(membar);
3506 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3607 lock->create_lock_counter(map()->jvms());
3608 increment_counter(lock->counter()->addr());
3609 }
3610 #endif
3611
3612 return flock;
3613 }
3614
3615
3616 //------------------------------shared_unlock----------------------------------
3617 // Emit unlocking code.
3618 void GraphKit::shared_unlock(Node* box, Node* obj) {
3619 // bci is either a monitorenter bc or InvocationEntryBci
3620 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3621 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3622
3623 if (stopped()) { // Dead monitor?
3624 map()->pop_monitor(); // Kill monitor from debug info
3625 return;
3626 }
3627
3628 // Memory barrier to avoid floating things down past the locked region
3629 insert_mem_bar(Op_MemBarReleaseLock);
3630
3631 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3632 UnlockNode *unlock = new UnlockNode(C, tf);
3633 #ifdef ASSERT
3634 unlock->set_dbg_jvms(sync_jvms());
3635 #endif
3636 uint raw_idx = Compile::AliasIdxRaw;
3637 unlock->init_req( TypeFunc::Control, control() );
3638 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3639 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3640 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3641 unlock->init_req( TypeFunc::ReturnAdr, top() );
3642
3643 unlock->init_req(TypeFunc::Parms + 0, obj);
3644 unlock->init_req(TypeFunc::Parms + 1, box);
3645 unlock = _gvn.transform(unlock)->as_Unlock();
3646
3647 Node* mem = reset_memory();
3648
3649 // unlock has no side-effects, sets few values
3650 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3651
3652 // Kill monitor from debug info
3653 map()->pop_monitor( );
3654 }
3655
3656 //-------------------------------get_layout_helper-----------------------------
3657 // If the given klass is a constant or known to be an array,
3658 // fetch the constant layout helper value into constant_value
3659 // and return null. Otherwise, load the non-constant
3660 // layout helper value, and return the node which represents it.
3661 // This two-faced routine is useful because allocation sites
3662 // almost always feature constant types.
3663 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3664 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3665 if (!StressReflectiveCode && klass_t != nullptr) {
3666 bool xklass = klass_t->klass_is_exact();
3667 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3668 jint lhelper;
3669 if (klass_t->isa_aryklassptr()) {
3670 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3671 if (is_reference_type(elem, true)) {
3672 elem = T_OBJECT;
3673 }
3674 lhelper = Klass::array_layout_helper(elem);
3675 } else {
3676 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3677 }
3678 if (lhelper != Klass::_lh_neutral_value) {
3679 constant_value = lhelper;
3680 return (Node*) nullptr;
3681 }
3682 }
3683 }
3684 constant_value = Klass::_lh_neutral_value; // put in a known value
3685 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
3686 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3687 }
3688
3689 // We just put in an allocate/initialize with a big raw-memory effect.
3690 // Hook selected additional alias categories on the initialization.
3691 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3692 MergeMemNode* init_in_merge,
3693 Node* init_out_raw) {
3694 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3695 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3696
3697 Node* prevmem = kit.memory(alias_idx);
3698 init_in_merge->set_memory_at(alias_idx, prevmem);
3699 kit.set_memory(init_out_raw, alias_idx);
3700 }
3701
3702 //---------------------------set_output_for_allocation-------------------------
3703 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3704 const TypeOopPtr* oop_type,
3705 bool deoptimize_on_exception) {
3706 int rawidx = Compile::AliasIdxRaw;
3707 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3708 add_safepoint_edges(alloc);
3709 Node* allocx = _gvn.transform(alloc);
3710 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3711 // create memory projection for i_o
3712 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3713 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3714
3715 // create a memory projection as for the normal control path
3716 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3717 set_memory(malloc, rawidx);
3718
3719 // a normal slow-call doesn't change i_o, but an allocation does
3720 // we create a separate i_o projection for the normal control path
3721 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3722 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3723
3724 // put in an initialization barrier
3725 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3726 rawoop)->as_Initialize();
3727 assert(alloc->initialization() == init, "2-way macro link must work");
3728 assert(init ->allocation() == alloc, "2-way macro link must work");
3729 {
3730 // Extract memory strands which may participate in the new object's
3731 // initialization, and source them from the new InitializeNode.
3732 // This will allow us to observe initializations when they occur,
3733 // and link them properly (as a group) to the InitializeNode.
3734 assert(init->in(InitializeNode::Memory) == malloc, "");
3735 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3736 init->set_req(InitializeNode::Memory, minit_in);
3737 record_for_igvn(minit_in); // fold it up later, if possible
3738 Node* minit_out = memory(rawidx);
3739 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3740 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
3741 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
3742 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
3743 // multiple projections as a result.
3744 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
3745 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
3746 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
3747 if (oop_type->isa_aryptr()) {
3748 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3749 int elemidx = C->get_alias_index(telemref);
3750 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(elemidx))));
3751 } else if (oop_type->isa_instptr()) {
3752 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3753 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3754 ciField* field = ik->nonstatic_field_at(i);
3755 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3756 continue; // do not bother to track really large numbers of fields
3757 // Find (or create) the alias category for this field:
3758 int fieldidx = C->alias_type(field)->index();
3759 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
3760 }
3761 }
3762 }
3763
3764 // Cast raw oop to the real thing...
3765 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3766 javaoop = _gvn.transform(javaoop);
3767 C->set_recent_alloc(control(), javaoop);
3768 assert(just_allocated_object(control()) == javaoop, "just allocated");
3769
3770 #ifdef ASSERT
3782 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3783 }
3784 }
3785 #endif //ASSERT
3786
3787 return javaoop;
3788 }
3789
3790 //---------------------------new_instance--------------------------------------
3791 // This routine takes a klass_node which may be constant (for a static type)
3792 // or may be non-constant (for reflective code). It will work equally well
3793 // for either, and the graph will fold nicely if the optimizer later reduces
3794 // the type to a constant.
3795 // The optional arguments are for specialized use by intrinsics:
3796 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3797 // - If 'return_size_val', report the total object size to the caller.
3798 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3799 Node* GraphKit::new_instance(Node* klass_node,
3800 Node* extra_slow_test,
3801 Node* *return_size_val,
3802 bool deoptimize_on_exception) {
3803 // Compute size in doublewords
3804 // The size is always an integral number of doublewords, represented
3805 // as a positive bytewise size stored in the klass's layout_helper.
3806 // The layout_helper also encodes (in a low bit) the need for a slow path.
3807 jint layout_con = Klass::_lh_neutral_value;
3808 Node* layout_val = get_layout_helper(klass_node, layout_con);
3809 int layout_is_con = (layout_val == nullptr);
3810
3811 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3812 // Generate the initial go-slow test. It's either ALWAYS (return a
3813 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3814 // case) a computed value derived from the layout_helper.
3815 Node* initial_slow_test = nullptr;
3816 if (layout_is_con) {
3817 assert(!StressReflectiveCode, "stress mode does not use these paths");
3818 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3819 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3820 } else { // reflective case
3821 // This reflective path is used by Unsafe.allocateInstance.
3822 // (It may be stress-tested by specifying StressReflectiveCode.)
3823 // Basically, we want to get into the VM is there's an illegal argument.
3824 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3825 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3826 if (extra_slow_test != intcon(0)) {
3827 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3828 }
3829 // (Macro-expander will further convert this to a Bool, if necessary.)
3835 if (layout_is_con) {
3836 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3837 } else { // reflective case
3838 // This reflective path is used by clone and Unsafe.allocateInstance.
3839 size = ConvI2X(layout_val);
3840
3841 // Clear the low bits to extract layout_helper_size_in_bytes:
3842 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3843 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3844 size = _gvn.transform( new AndXNode(size, mask) );
3845 }
3846 if (return_size_val != nullptr) {
3847 (*return_size_val) = size;
3848 }
3849
3850 // This is a precise notnull oop of the klass.
3851 // (Actually, it need not be precise if this is a reflective allocation.)
3852 // It's what we cast the result to.
3853 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3854 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3855 const TypeOopPtr* oop_type = tklass->as_instance_type();
3856
3857 // Now generate allocation code
3858
3859 // The entire memory state is needed for slow path of the allocation
3860 // since GC and deoptimization can happened.
3861 Node *mem = reset_memory();
3862 set_all_memory(mem); // Create new memory state
3863
3864 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3865 control(), mem, i_o(),
3866 size, klass_node,
3867 initial_slow_test);
3868
3869 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3870 }
3871
3872 //-------------------------------new_array-------------------------------------
3873 // helper for both newarray and anewarray
3874 // The 'length' parameter is (obviously) the length of the array.
3875 // The optional arguments are for specialized use by intrinsics:
3876 // - If 'return_size_val', report the non-padded array size (sum of header size
3877 // and array body) to the caller.
3878 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3879 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3880 Node* length, // number of array elements
3881 int nargs, // number of arguments to push back for uncommon trap
3882 Node* *return_size_val,
3883 bool deoptimize_on_exception) {
3884 jint layout_con = Klass::_lh_neutral_value;
3885 Node* layout_val = get_layout_helper(klass_node, layout_con);
3886 int layout_is_con = (layout_val == nullptr);
3887
3888 if (!layout_is_con && !StressReflectiveCode &&
3889 !too_many_traps(Deoptimization::Reason_class_check)) {
3890 // This is a reflective array creation site.
3891 // Optimistically assume that it is a subtype of Object[],
3892 // so that we can fold up all the address arithmetic.
3893 layout_con = Klass::array_layout_helper(T_OBJECT);
3894 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3895 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3896 { BuildCutout unless(this, bol_lh, PROB_MAX);
3897 inc_sp(nargs);
3898 uncommon_trap(Deoptimization::Reason_class_check,
3899 Deoptimization::Action_maybe_recompile);
3900 }
3901 layout_val = nullptr;
3902 layout_is_con = true;
3903 }
3904
3905 // Generate the initial go-slow test. Make sure we do not overflow
3906 // if length is huge (near 2Gig) or negative! We do not need
3907 // exact double-words here, just a close approximation of needed
3908 // double-words. We can't add any offset or rounding bits, lest we
3909 // take a size -1 of bytes and make it positive. Use an unsigned
3910 // compare, so negative sizes look hugely positive.
3911 int fast_size_limit = FastAllocateSizeLimit;
3912 if (layout_is_con) {
3913 assert(!StressReflectiveCode, "stress mode does not use these paths");
3914 // Increase the size limit if we have exact knowledge of array type.
3915 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3916 assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3917 "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3918 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3919 }
3920
3921 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3922 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3923
3924 // --- Size Computation ---
3925 // array_size = round_to_heap(array_header + (length << elem_shift));
3926 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3927 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3928 // The rounding mask is strength-reduced, if possible.
3929 int round_mask = MinObjAlignmentInBytes - 1;
3930 Node* header_size = nullptr;
3931 // (T_BYTE has the weakest alignment and size restrictions...)
3932 if (layout_is_con) {
3933 int hsize = Klass::layout_helper_header_size(layout_con);
3934 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3935 if ((round_mask & ~right_n_bits(eshift)) == 0)
3936 round_mask = 0; // strength-reduce it if it goes away completely
3937 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3938 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3939 assert(header_size_min <= hsize, "generic minimum is smallest");
3940 header_size = intcon(hsize);
3941 } else {
3942 Node* hss = intcon(Klass::_lh_header_size_shift);
3943 Node* hsm = intcon(Klass::_lh_header_size_mask);
3944 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3945 header_size = _gvn.transform(new AndINode(header_size, hsm));
3946 }
3947
3948 Node* elem_shift = nullptr;
3949 if (layout_is_con) {
3950 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3951 if (eshift != 0)
3952 elem_shift = intcon(eshift);
3953 } else {
3954 // There is no need to mask or shift this value.
3955 // The semantics of LShiftINode include an implicit mask to 0x1F.
3956 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3957 elem_shift = layout_val;
4006 }
4007 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4008
4009 if (return_size_val != nullptr) {
4010 // This is the size
4011 (*return_size_val) = non_rounded_size;
4012 }
4013
4014 Node* size = non_rounded_size;
4015 if (round_mask != 0) {
4016 Node* mask1 = MakeConX(round_mask);
4017 size = _gvn.transform(new AddXNode(size, mask1));
4018 Node* mask2 = MakeConX(~round_mask);
4019 size = _gvn.transform(new AndXNode(size, mask2));
4020 }
4021 // else if round_mask == 0, the size computation is self-rounding
4022
4023 // Now generate allocation code
4024
4025 // The entire memory state is needed for slow path of the allocation
4026 // since GC and deoptimization can happened.
4027 Node *mem = reset_memory();
4028 set_all_memory(mem); // Create new memory state
4029
4030 if (initial_slow_test->is_Bool()) {
4031 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4032 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4033 }
4034
4035 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
4036 Node* valid_length_test = _gvn.intcon(1);
4037 if (ary_type->isa_aryptr()) {
4038 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4039 jint max = TypeAryPtr::max_array_length(bt);
4040 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4041 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4042 }
4043
4044 // Create the AllocateArrayNode and its result projections
4045 AllocateArrayNode* alloc
4046 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4047 control(), mem, i_o(),
4048 size, klass_node,
4049 initial_slow_test,
4050 length, valid_length_test);
4051
4052 // Cast to correct type. Note that the klass_node may be constant or not,
4053 // and in the latter case the actual array type will be inexact also.
4054 // (This happens via a non-constant argument to inline_native_newArray.)
4055 // In any case, the value of klass_node provides the desired array type.
4056 const TypeInt* length_type = _gvn.find_int_type(length);
4057 if (ary_type->isa_aryptr() && length_type != nullptr) {
4058 // Try to get a better type than POS for the size
4059 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4060 }
4061
4062 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4063
4064 array_ideal_length(alloc, ary_type, true);
4065 return javaoop;
4066 }
4067
4068 // The following "Ideal_foo" functions are placed here because they recognize
4069 // the graph shapes created by the functions immediately above.
4070
4071 //---------------------------Ideal_allocation----------------------------------
4166 void GraphKit::add_parse_predicates(int nargs) {
4167 if (ShortRunningLongLoop) {
4168 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4169 // walking up from the loop.
4170 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4171 }
4172 if (UseLoopPredicate) {
4173 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4174 if (UseProfiledLoopPredicate) {
4175 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4176 }
4177 }
4178 if (UseAutoVectorizationPredicate) {
4179 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4180 }
4181 // Loop Limit Check Predicate should be near the loop.
4182 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4183 }
4184
4185 void GraphKit::sync_kit(IdealKit& ideal) {
4186 set_all_memory(ideal.merged_memory());
4187 set_i_o(ideal.i_o());
4188 set_control(ideal.ctrl());
4189 }
4190
4191 void GraphKit::final_sync(IdealKit& ideal) {
4192 // Final sync IdealKit and graphKit.
4193 sync_kit(ideal);
4194 }
4195
4196 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4197 Node* len = load_array_length(load_String_value(str, set_ctrl));
4198 Node* coder = load_String_coder(str, set_ctrl);
4199 // Divide length by 2 if coder is UTF16
4200 return _gvn.transform(new RShiftINode(len, coder));
4201 }
4202
4203 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4204 int value_offset = java_lang_String::value_offset();
4205 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4206 false, nullptr, 0);
4207 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4208 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::BotPTR,
4209 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4210 ciTypeArrayKlass::make(T_BYTE), true, 0);
4211 Node* p = basic_plus_adr(str, str, value_offset);
4212 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4213 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4214 return must_be_not_null(load, true);
4215 }
4216
4217 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4218 if (!CompactStrings) {
4219 return intcon(java_lang_String::CODER_UTF16);
4220 }
4221 int coder_offset = java_lang_String::coder_offset();
4222 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4223 false, nullptr, 0);
4224 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4225
4226 Node* p = basic_plus_adr(str, str, coder_offset);
4227 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4228 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4229 return load;
4230 }
4231
4232 void GraphKit::store_String_value(Node* str, Node* value) {
4233 int value_offset = java_lang_String::value_offset();
4234 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4235 false, nullptr, 0);
4236 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4237
4238 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4239 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4240 }
4241
4242 void GraphKit::store_String_coder(Node* str, Node* value) {
4243 int coder_offset = java_lang_String::coder_offset();
4244 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4245 false, nullptr, 0);
4246 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4247
4248 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4249 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4250 }
4251
4252 // If input and output memory types differ, capture the whole memory to preserve
4253 // the dependency between preceding and subsequent loads/stores.
4254 // For example, the following program:
4255 // StoreB
4256 // compress_string
4257 // LoadB
4258 // has this memory graph (use->def):
4259 // LoadB -> compress_string -> CharMem
4260 // ... -> StoreB -> ByteMem
4261 // The intrinsic hides the dependency between LoadB and StoreB, causing
4262 // the load to read from memory not containing the result of the StoreB.
4263 // The correct memory graph should look like this:
4264 // LoadB -> compress_string -> MergeMem -> StoreB
4265 Node* GraphKit::capture_memory(const TypePtr*& combined_type, const TypePtr* src_type, const TypePtr* dst_type) {
4368 i_char->init_req(2, AddI(i_char, intcon(2)));
4369
4370 set_control(IfFalse(iff));
4371 set_memory(st, TypeAryPtr::BYTES);
4372 }
4373
4374 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4375 if (!field->is_constant()) {
4376 return nullptr; // Field not marked as constant.
4377 }
4378 ciInstance* holder = nullptr;
4379 if (!field->is_static()) {
4380 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4381 if (const_oop != nullptr && const_oop->is_instance()) {
4382 holder = const_oop->as_instance();
4383 }
4384 }
4385 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4386 /*is_unsigned_load=*/false);
4387 if (con_type != nullptr) {
4388 return makecon(con_type);
4389 }
4390 return nullptr;
4391 }
4392
4393 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4394 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4395 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4396 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4397 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4398 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4399 return casted_obj;
4400 }
4401 return obj;
4402 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/graphKit.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/multnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/opcodes.hpp"
53 #include "opto/parse.hpp"
54 #include "opto/reachability.hpp"
55 #include "opto/rootnode.hpp"
56 #include "opto/runtime.hpp"
57 #include "opto/subtypenode.hpp"
58 #include "opto/type.hpp"
59 #include "runtime/arguments.hpp"
60 #include "runtime/deoptimization.hpp"
61 #include "runtime/sharedRuntime.hpp"
62 #include "runtime/stubRoutines.hpp"
63 #include "utilities/bitMap.inline.hpp"
64 #include "utilities/growableArray.hpp"
65 #include "utilities/powerOfTwo.hpp"
66
67 //----------------------------GraphKit-----------------------------------------
68 // Main utility constructor.
69 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
70 : Phase(Phase::Parser),
71 _env(C->env()),
72 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
73 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
74 {
75 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
76 _exceptions = jvms->map()->next_exception();
77 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
78 set_jvms(jvms);
79 #ifdef ASSERT
80 if (_gvn.is_IterGVN() != nullptr) {
81 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
82 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
83 _worklist_size = _gvn.C->igvn_worklist()->size();
84 }
85 #endif
86 }
87
88 // Private constructor for parser.
89 GraphKit::GraphKit()
90 : Phase(Phase::Parser),
91 _env(C->env()),
92 _gvn(*C->initial_gvn()),
93 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
94 {
95 _exceptions = nullptr;
96 set_map(nullptr);
97 DEBUG_ONLY(_sp = -99);
98 DEBUG_ONLY(set_bci(-99));
99 }
100
101 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
102 : Phase(Phase::Parser),
103 _env(C->env()),
104 _gvn(igvn),
105 _exceptions(nullptr),
106 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
107 assert(igvn.delay_transform(), "must delay transformation during macro expansion");
108 assert(sft->next_exception() == nullptr, "must not have a pending exception");
109 JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
110 SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
111 for (uint i = 0; i < sft->req(); i++) {
112 cloned_map->init_req(i, sft->in(i));
113 }
114 igvn.record_for_igvn(cloned_map);
115 for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
116 current->set_map(cloned_map);
117 }
118 set_jvms(cloned_jvms);
119 set_all_memory(cloned_map->memory());
120 }
121
122 //---------------------------clean_stack---------------------------------------
123 // Clear away rubbish from the stack area of the JVM state.
124 // This destroys any arguments that may be waiting on the stack.
125 void GraphKit::clean_stack(int from_sp) {
126 SafePointNode* map = this->map();
127 JVMState* jvms = this->jvms();
128 int stk_size = jvms->stk_size();
129 int stkoff = jvms->stkoff();
130 Node* top = this->top();
131 for (int i = from_sp; i < stk_size; i++) {
132 if (map->in(stkoff + i) != top) {
133 map->set_req(stkoff + i, top);
134 }
135 }
136 }
137
138
139 //--------------------------------sync_jvms-----------------------------------
140 // Make sure our current jvms agrees with our parse state.
939 if (PrintMiscellaneous && (Verbose || WizardMode)) {
940 tty->print_cr("Zombie local %d: ", local);
941 jvms->dump();
942 }
943 return false;
944 }
945 }
946 }
947 return true;
948 }
949
950 #endif //ASSERT
951
952 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
953 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
954 ciMethod* cur_method = jvms->method();
955 int cur_bci = jvms->bci();
956 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
957 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
958 return Interpreter::bytecode_should_reexecute(code) ||
959 (is_anewarray && (code == Bytecodes::_multianewarray));
960 // Reexecute _multianewarray bytecode which was replaced with
961 // sequence of [a]newarray. See Parse::do_multianewarray().
962 //
963 // Note: interpreter should not have it set since this optimization
964 // is limited by dimensions and guarded by flag so in some cases
965 // multianewarray() runtime calls will be generated and
966 // the bytecode should not be reexecutes (stack will not be reset).
967 } else {
968 return false;
969 }
970 }
971
972 // Helper function for adding JVMState and debug information to node
973 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
974 // Add the safepoint edges to the call (or other safepoint).
975
976 // Make sure dead locals are set to top. This
977 // should help register allocation time and cut down on the size
978 // of the deoptimization information.
979 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1055 uint p = debug_start; // walks forward in [debug_start, debug_end)
1056 uint j, k, l;
1057 SafePointNode* in_map = in_jvms->map();
1058 out_jvms->set_map(call);
1059
1060 if (can_prune_locals) {
1061 assert(in_jvms->method() == out_jvms->method(), "sanity");
1062 // If the current throw can reach an exception handler in this JVMS,
1063 // then we must keep everything live that can reach that handler.
1064 // As a quick and dirty approximation, we look for any handlers at all.
1065 if (in_jvms->method()->has_exception_handlers()) {
1066 can_prune_locals = false;
1067 }
1068 }
1069
1070 // Add the Locals
1071 k = in_jvms->locoff();
1072 l = in_jvms->loc_size();
1073 out_jvms->set_locoff(p);
1074 if (!can_prune_locals) {
1075 for (j = 0; j < l; j++) {
1076 call->set_req(p++, in_map->in(k + j));
1077 }
1078 } else {
1079 p += l; // already set to top above by add_req_batch
1080 }
1081
1082 // Add the Expression Stack
1083 k = in_jvms->stkoff();
1084 l = in_jvms->sp();
1085 out_jvms->set_stkoff(p);
1086 if (!can_prune_locals) {
1087 for (j = 0; j < l; j++) {
1088 call->set_req(p++, in_map->in(k + j));
1089 }
1090 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1091 // Divide stack into {S0,...,S1}, where S0 is set to top.
1092 uint s1 = stack_slots_not_pruned;
1093 stack_slots_not_pruned = 0; // for next iteration
1094 if (s1 > l) s1 = l;
1095 uint s0 = l - s1;
1096 p += s0; // skip the tops preinstalled by add_req_batch
1097 for (j = s0; j < l; j++)
1098 call->set_req(p++, in_map->in(k+j));
1099 } else {
1100 p += l; // already set to top above by add_req_batch
1101 }
1102
1103 // Add the Monitors
1104 k = in_jvms->monoff();
1105 l = in_jvms->mon_size();
1106 out_jvms->set_monoff(p);
1107 for (j = 0; j < l; j++)
1108 call->set_req(p++, in_map->in(k+j));
1109
1297 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1298 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1299 return _gvn.transform( new AndLNode(conv, mask) );
1300 }
1301
1302 Node* GraphKit::ConvL2I(Node* offset) {
1303 // short-circuit a common case
1304 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1305 if (offset_con != (jlong)Type::OffsetBot) {
1306 return intcon((int) offset_con);
1307 }
1308 return _gvn.transform( new ConvL2INode(offset));
1309 }
1310
1311 //-------------------------load_object_klass-----------------------------------
1312 Node* GraphKit::load_object_klass(Node* obj) {
1313 // Special-case a fresh allocation to avoid building nodes:
1314 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1315 if (akls != nullptr) return akls;
1316 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1317 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1318 }
1319
1320 //-------------------------load_array_length-----------------------------------
1321 Node* GraphKit::load_array_length(Node* array) {
1322 // Special-case a fresh allocation to avoid building nodes:
1323 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1324 Node *alen;
1325 if (alloc == nullptr) {
1326 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1327 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1328 } else {
1329 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1330 }
1331 return alen;
1332 }
1333
1334 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1335 const TypeOopPtr* oop_type,
1336 bool replace_length_in_map) {
1337 Node* length = alloc->Ideal_length();
1346 replace_in_map(length, ccast);
1347 }
1348 return ccast;
1349 }
1350 }
1351 return length;
1352 }
1353
1354 //------------------------------do_null_check----------------------------------
1355 // Helper function to do a null pointer check. Returned value is
1356 // the incoming address with null casted away. You are allowed to use the
1357 // not-null value only if you are control dependent on the test.
1358 #ifndef PRODUCT
1359 extern uint explicit_null_checks_inserted,
1360 explicit_null_checks_elided;
1361 #endif
1362 Node* GraphKit::null_check_common(Node* value, BasicType type,
1363 // optional arguments for variations:
1364 bool assert_null,
1365 Node* *null_control,
1366 bool speculative,
1367 bool null_marker_check) {
1368 assert(!assert_null || null_control == nullptr, "not both at once");
1369 if (stopped()) return top();
1370 NOT_PRODUCT(explicit_null_checks_inserted++);
1371
1372 if (value->is_InlineType()) {
1373 // Null checking a scalarized but nullable inline type. Check the null marker
1374 // input instead of the oop input to avoid keeping buffer allocations alive.
1375 null_check_common(value->as_InlineType()->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1376 if (stopped()) {
1377 return top();
1378 }
1379 if (assert_null) {
1380 replace_in_map(value, null());
1381 return null();
1382 }
1383 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1384 return cast_not_null(value, do_replace_in_map);
1385 }
1386
1387 // Construct null check
1388 Node *chk = nullptr;
1389 switch(type) {
1390 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1391 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1392 case T_ARRAY : // fall through
1393 type = T_OBJECT; // simplify further tests
1394 case T_OBJECT : {
1395 const Type *t = _gvn.type( value );
1396
1397 const TypeOopPtr* tp = t->isa_oopptr();
1398 if (tp != nullptr && !tp->is_loaded()
1399 // Only for do_null_check, not any of its siblings:
1400 && !assert_null && null_control == nullptr) {
1401 // Usually, any field access or invocation on an unloaded oop type
1402 // will simply fail to link, since the statically linked class is
1403 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1404 // the static class is loaded but the sharper oop type is not.
1405 // Rather than checking for this obscure case in lots of places,
1406 // we simply observe that a null check on an unloaded class
1470 }
1471 Node *oldcontrol = control();
1472 set_control(cfg);
1473 Node *res = cast_not_null(value);
1474 set_control(oldcontrol);
1475 NOT_PRODUCT(explicit_null_checks_elided++);
1476 return res;
1477 }
1478 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1479 if (cfg == nullptr) break; // Quit at region nodes
1480 depth++;
1481 }
1482 }
1483
1484 //-----------
1485 // Branch to failure if null
1486 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1487 Deoptimization::DeoptReason reason;
1488 if (assert_null) {
1489 reason = Deoptimization::reason_null_assert(speculative);
1490 } else if (type == T_OBJECT || null_marker_check) {
1491 reason = Deoptimization::reason_null_check(speculative);
1492 } else {
1493 reason = Deoptimization::Reason_div0_check;
1494 }
1495 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1496 // ciMethodData::has_trap_at will return a conservative -1 if any
1497 // must-be-null assertion has failed. This could cause performance
1498 // problems for a method after its first do_null_assert failure.
1499 // Consider using 'Reason_class_check' instead?
1500
1501 // To cause an implicit null check, we set the not-null probability
1502 // to the maximum (PROB_MAX). For an explicit check the probability
1503 // is set to a smaller value.
1504 if (null_control != nullptr || too_many_traps(reason)) {
1505 // probability is less likely
1506 ok_prob = PROB_LIKELY_MAG(3);
1507 } else if (!assert_null &&
1508 (ImplicitNullCheckThreshold > 0) &&
1509 method() != nullptr &&
1510 (method()->method_data()->trap_count(reason)
1544 }
1545
1546 if (assert_null) {
1547 // Cast obj to null on this path.
1548 replace_in_map(value, zerocon(type));
1549 return zerocon(type);
1550 }
1551
1552 // Cast obj to not-null on this path, if there is no null_control.
1553 // (If there is a null_control, a non-null value may come back to haunt us.)
1554 if (type == T_OBJECT) {
1555 Node* cast = cast_not_null(value, false);
1556 if (null_control == nullptr || (*null_control) == top())
1557 replace_in_map(value, cast);
1558 value = cast;
1559 }
1560
1561 return value;
1562 }
1563
1564 //------------------------------cast_not_null----------------------------------
1565 // Cast obj to not-null on this path
1566 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1567 const Type* t = _gvn.type(obj);
1568 const Type* t_not_null = t->join_speculative(TypePtr::NOTNULL);
1569 if (t == t_not_null) {
1570 return obj;
1571 }
1572
1573 Node* cast = new CastPPNode(control(), obj, t_not_null);
1574 cast = _gvn.transform(cast);
1575
1576 // Scan for instances of 'obj' in the current JVM mapping.
1577 // These instances are known to be not-null after the test.
1578 if (do_replace_in_map) {
1579 replace_in_map(obj, cast);
1580 }
1581 return cast;
1582 }
1583
1584 // Sometimes in intrinsics, we implicitly know an object is not null
1585 // (there's no actual null check) so we can cast it to not null. In
1586 // the course of optimizations, the input to the cast can become null.
1587 // In that case that data path will die and we need the control path
1588 // to become dead as well to keep the graph consistent. So we have to
1589 // add a check for null for which one branch can't be taken. It uses
1590 // an OpaqueConstantBool node that will cause the check to be removed after loop
1591 // opts so the test goes away and the compiled code doesn't execute a
1592 // useless check.
1593 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1594 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1595 return value;
1596 }
1597 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1598 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1599 Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1600 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1601 _gvn.set_type(iff, iff->Value(&_gvn));
1640 //=============================================================================
1641 //--------------------------------memory---------------------------------------
1642 Node* GraphKit::memory(uint alias_idx) {
1643 MergeMemNode* mem = merged_memory();
1644 Node* p = mem->memory_at(alias_idx);
1645 assert(p != mem->empty_memory(), "empty");
1646 _gvn.set_type(p, Type::MEMORY); // must be mapped
1647 return p;
1648 }
1649
1650 //-----------------------------reset_memory------------------------------------
1651 Node* GraphKit::reset_memory() {
1652 Node* mem = map()->memory();
1653 // do not use this node for any more parsing!
1654 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1655 return _gvn.transform( mem );
1656 }
1657
1658 //------------------------------set_all_memory---------------------------------
1659 void GraphKit::set_all_memory(Node* newmem) {
1660 // The 2 cases are semantically equivalent
1661 MergeMemNode* mergemem;
1662 if (_gvn.is_IterGVN()) {
1663 // During IGVN, create a more predictable pattern so it is easier to verify that the GraphKit
1664 // does not modify memory
1665 mergemem = MergeMemNode::make(C->top());
1666 mergemem->set_base_memory(newmem);
1667 } else {
1668 // During parsing, be a little more aggressive so that GVN can fold accesses more easily
1669 mergemem = MergeMemNode::make(newmem);
1670 }
1671 _gvn.set_type_bottom(mergemem);
1672 record_for_igvn(mergemem);
1673 map()->set_memory(mergemem);
1674 }
1675
1676 //------------------------------set_all_memory_call----------------------------
1677 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1678 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1679 set_all_memory(newmem);
1680 }
1681
1682 //=============================================================================
1683 //
1684 // parser factory methods for MemNodes
1685 //
1686 // These are layered on top of the factory methods in LoadNode and StoreNode,
1687 // and integrate with the parser's memory state and _gvn engine.
1688 //
1689
1690 // factory methods in "int adr_idx"
1691 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1692 MemNode::MemOrd mo,
1693 LoadNode::ControlDependency control_dependency,
1694 bool require_atomic_access,
1695 bool unaligned,
1696 bool mismatched,
1697 bool unsafe,
1698 uint8_t barrier_data) {
1699 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1700 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1701 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1702 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1703 Node* mem = memory(adr_idx);
1704 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1705 ld = _gvn.transform(ld);
1706
1707 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1708 // Improve graph before escape analysis and boxing elimination.
1709 record_for_igvn(ld);
1710 if (ld->is_DecodeN()) {
1711 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1712 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1713 // a Phi). Recording such cases is still perfectly sound, but may be
1714 // unnecessary and result in some minor IGVN overhead.
1715 record_for_igvn(ld->in(1));
1716 }
1717 }
1718 return ld;
1719 }
1720
1721 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1722 MemNode::MemOrd mo,
1723 bool require_atomic_access,
1724 bool unaligned,
1725 bool mismatched,
1726 bool unsafe,
1740 if (unsafe) {
1741 st->as_Store()->set_unsafe_access();
1742 }
1743 st->as_Store()->set_barrier_data(barrier_data);
1744 st = _gvn.transform(st);
1745 set_memory(st, adr_idx);
1746 // Back-to-back stores can only remove intermediate store with DU info
1747 // so push on worklist for optimizer.
1748 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1749 record_for_igvn(st);
1750
1751 return st;
1752 }
1753
1754 Node* GraphKit::access_store_at(Node* obj,
1755 Node* adr,
1756 const TypePtr* adr_type,
1757 Node* val,
1758 const Type* val_type,
1759 BasicType bt,
1760 DecoratorSet decorators,
1761 bool safe_for_replace,
1762 const InlineTypeNode* vt) {
1763 // Transformation of a value which could be null pointer (CastPP #null)
1764 // could be delayed during Parse (for example, in adjust_map_after_if()).
1765 // Execute transformation here to avoid barrier generation in such case.
1766 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1767 val = _gvn.makecon(TypePtr::NULL_PTR);
1768 }
1769
1770 if (stopped()) {
1771 return top(); // Dead path ?
1772 }
1773
1774 assert(val != nullptr, "not dead path");
1775 if (val->is_InlineType()) {
1776 // Store to non-flat field. Buffer the inline type and make sure
1777 // the store is re-executed if the allocation triggers deoptimization.
1778 PreserveReexecuteState preexecs(this);
1779 jvms()->set_should_reexecute(true);
1780 val = val->as_InlineType()->buffer(this, safe_for_replace);
1781 }
1782
1783 C2AccessValuePtr addr(adr, adr_type);
1784 C2AccessValue value(val, val_type);
1785 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1786 if (access.is_raw()) {
1787 return _barrier_set->BarrierSetC2::store_at(access, value);
1788 } else {
1789 return _barrier_set->store_at(access, value);
1790 }
1791 }
1792
1793 Node* GraphKit::access_load_at(Node* obj, // containing obj
1794 Node* adr, // actual address to store val at
1795 const TypePtr* adr_type,
1796 const Type* val_type,
1797 BasicType bt,
1798 DecoratorSet decorators,
1799 Node* ctl) {
1800 if (stopped()) {
1801 return top(); // Dead path ?
1802 }
1803
1804 SavedState old_state(this);
1805 C2AccessValuePtr addr(adr, adr_type);
1806 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1807 Node* load;
1808 if (access.is_raw()) {
1809 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1810 } else {
1811 load = _barrier_set->load_at(access, val_type);
1812 }
1813
1814 // Restore the previous state only if the load got folded to a constant
1815 // and we can discard any barriers that might have been added.
1816 if (load == nullptr || !load->is_Con()) {
1817 old_state.discard();
1818 }
1819 return load;
1820 }
1821
1822 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1823 const Type* val_type,
1824 BasicType bt,
1825 DecoratorSet decorators) {
1826 if (stopped()) {
1908 Node* new_val,
1909 const Type* value_type,
1910 BasicType bt,
1911 DecoratorSet decorators) {
1912 C2AccessValuePtr addr(adr, adr_type);
1913 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1914 if (access.is_raw()) {
1915 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1916 } else {
1917 return _barrier_set->atomic_add_at(access, new_val, value_type);
1918 }
1919 }
1920
1921 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1922 return _barrier_set->clone(this, src, dst, size, is_array);
1923 }
1924
1925 //-------------------------array_element_address-------------------------
1926 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1927 const TypeInt* sizetype, Node* ctrl) {
1928 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1929 uint shift;
1930 uint header;
1931 if (arytype->is_flat() && arytype->klass_is_exact()) {
1932 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1933 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1934 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1935 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1936 // though we don't need the address node in this case and throw it away again.
1937 shift = arytype->flat_log_elem_size();
1938 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1939 } else {
1940 shift = exact_log2(type2aelembytes(elembt));
1941 header = arrayOopDesc::base_offset_in_bytes(elembt);
1942 }
1943
1944 // short-circuit a common case (saves lots of confusing waste motion)
1945 jint idx_con = find_int_con(idx, -1);
1946 if (idx_con >= 0) {
1947 intptr_t offset = header + ((intptr_t)idx_con << shift);
1948 return basic_plus_adr(ary, offset);
1949 }
1950
1951 // must be correct type for alignment purposes
1952 Node* base = basic_plus_adr(ary, header);
1953 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1954 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1955 return basic_plus_adr(ary, base, scale);
1956 }
1957
1958 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1959 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1960 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1961 return cast_to_flat_array_exact(array, elem_vk, true, false);
1962 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1963 return cast_to_flat_array_exact(array, elem_vk, true, true);
1964 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1965 return cast_to_flat_array_exact(array, elem_vk, false, true);
1966 }
1967
1968 bool is_null_free = false;
1969 if (!elem_vk->has_nullable_atomic_layout()) {
1970 // Element does not have a nullable flat layout, cannot be nullable
1971 is_null_free = true;
1972 }
1973
1974 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1975 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1976 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1977 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1978 }
1979
1980 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1981 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1982 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1983 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1984 assert(arytype->klass_is_exact(), "inconsistency");
1985 assert(arytype->is_flat(), "inconsistency");
1986 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1987 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1988 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1989 }
1990
1991 //-------------------------load_array_element-------------------------
1992 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1993 const Type* elemtype = arytype->elem();
1994 BasicType elembt = elemtype->array_element_basic_type();
1995 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1996 if (elembt == T_NARROWOOP) {
1997 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1998 }
1999 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
2000 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
2001 return ld;
2002 }
2003
2004 //-------------------------set_arguments_for_java_call-------------------------
2005 // Arguments (pre-popped from the stack) are taken from the JVMS.
2006 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
2007 PreserveReexecuteState preexecs(this);
2008 if (Arguments::is_valhalla_enabled()) {
2009 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
2010 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
2011 jvms()->set_should_reexecute(true);
2012 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
2013 inc_sp(arg_size);
2014 }
2015 // Add the call arguments
2016 const TypeTuple* domain = call->tf()->domain_sig();
2017 uint nargs = domain->cnt();
2018 int arg_num = 0;
2019 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
2020 uint arg_idx = i - TypeFunc::Parms;
2021 Node* arg = argument(arg_idx);
2022 const Type* t = domain->field_at(i);
2023 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
2024 // We don't pass inline type arguments by reference but instead pass each field of the inline type
2025 if (!arg->is_InlineType()) {
2026 // There are 2 cases in which the argument has not been scalarized
2027 if (_gvn.type(arg)->is_zero_type()) {
2028 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
2029 } else {
2030 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
2031 // receiver is a non-scalarized oop. CHA or IGVN might then prove that the receiver
2032 // type must be an exact value class. The method is devirtualized, and replaced with
2033 // a direct call with a scalarized receiver instead.
2034 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
2035 assert(call->is_optimized_virtual(), "must be during devirtualization of calls");
2036 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
2037 }
2038 }
2039 InlineTypeNode* vt = arg->as_InlineType();
2040 vt->pass_fields(this, call, idx, true, !t->maybe_null(), true);
2041 // If an inline type argument is passed as fields, attach the Method* to the call site
2042 // to be able to access the extended signature later via attached_method_before_pc().
2043 // For example, see CompiledMethod::preserve_callee_argument_oops().
2044 call->set_override_symbolic_info(true);
2045 // Register a calling convention dependency on the callee method to make sure that this method is deoptimized and
2046 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
2047 C->dependencies()->assert_mismatch_calling_convention(call->method());
2048 arg_num++;
2049 continue;
2050 } else if (arg->is_InlineType()) {
2051 // Pass inline type argument via oop to callee
2052 arg = arg->as_InlineType()->buffer(this, true);
2053 }
2054 if (t != Type::HALF) {
2055 arg_num++;
2056 }
2057 call->init_req(idx++, arg);
2058 }
2059 }
2060
2061 //---------------------------set_edges_for_java_call---------------------------
2062 // Connect a newly created call into the current JVMS.
2063 // A return value node (if any) is returned from set_edges_for_java_call.
2064 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2065
2066 // Add the predefined inputs:
2067 call->init_req( TypeFunc::Control, control() );
2068 call->init_req( TypeFunc::I_O , i_o() );
2069 call->init_req( TypeFunc::Memory , reset_memory() );
2070 call->init_req( TypeFunc::FramePtr, frameptr() );
2071 call->init_req( TypeFunc::ReturnAdr, top() );
2072
2073 add_safepoint_edges(call, must_throw);
2074
2075 Node* xcall = _gvn.transform(call);
2076
2077 if (xcall == top()) {
2078 set_control(top());
2079 return;
2080 }
2081 assert(xcall == call, "call identity is stable");
2082
2083 // Re-use the current map to produce the result.
2084
2085 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2086 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2087 set_all_memory_call(xcall, separate_io_proj);
2088
2089 //return xcall; // no need, caller already has it
2090 }
2091
2092 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2093 if (stopped()) return top(); // maybe the call folded up?
2094
2095 // Note: Since any out-of-line call can produce an exception,
2096 // we always insert an I_O projection from the call into the result.
2097
2098 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2099
2100 if (separate_io_proj) {
2101 // The caller requested separate projections be used by the fall
2102 // through and exceptional paths, so replace the projections for
2103 // the fall through path.
2104 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2105 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2106 }
2107
2108 // Capture the return value, if any.
2109 Node* ret;
2110 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2111 ret = top();
2112 } else if (call->tf()->returns_inline_type_as_fields()) {
2113 // Return of multiple values (inline type fields): we create a
2114 // InlineType node, each field is a projection from the call.
2115 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2116 uint base_input = TypeFunc::Parms;
2117 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2118 // If we run out of registers to store the null marker, we need to reserve an extra
2119 // slot to store it on the stack. Unfortunately, we only know if stack slot is needed
2120 // when matching the call (see Matcher::return_values_mask), so we are conservative here.
2121 C->set_needs_nm_slot(true);
2122 } else {
2123 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2124 ciType* t = call->method()->return_type();
2125 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2126 // The return type is unloaded but the callee might later be C2 compiled and then return
2127 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2128 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2129 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2130 IdealKit ideal(this);
2131 IdealVariable res(ideal);
2132 ideal.declarations_done();
2133 // Change return type of call to scalarized return
2134 const TypeFunc* tf = call->_tf;
2135 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2136 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2137 call->_tf = new_tf;
2138 _gvn.set_type(call, call->Value(&_gvn));
2139 _gvn.set_type(ret, ret->Value(&_gvn));
2140 // Don't add store to buffer call if we are strength reducing
2141 if (!C->strength_reduction()) {
2142 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2143 // Return value is null
2144 ideal.set(res, makecon(TypePtr::NULL_PTR));
2145 } ideal.else_(); {
2146 // Return value is non-null
2147 sync_kit(ideal);
2148
2149 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2150 OptoRuntime::store_inline_type_fields_Type(),
2151 SharedRuntime::store_inline_type_fields_to_buf_entry(),
2152 nullptr, TypePtr::BOTTOM, ret);
2153
2154 // We don't know how many values are returned. This assumes the
2155 // worst case, that all available registers are used.
2156 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2157 if (domain->field_at(i) == Type::HALF) {
2158 store_to_buf_call->init_req(i, top());
2159 continue;
2160 }
2161 Node* proj =_gvn.transform(new ProjNode(call, i));
2162 store_to_buf_call->init_req(i, proj);
2163 }
2164 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2165
2166 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2167 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2168 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2169
2170 ideal.set(res, buf);
2171 ideal.sync_kit(this);
2172 } ideal.end_if();
2173 } else {
2174 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2175 // Will be rewired later in replace_call().
2176 _gvn.transform(new ProjNode(call, i));
2177 }
2178 ideal.set(res, ret);
2179 }
2180 sync_kit(ideal);
2181 ret = _gvn.transform(ideal.value(res));
2182 } else if (!call->method()->return_value_is_larval() && _gvn.type(ret)->is_inlinetypeptr()) {
2183 // In Parse::do_call we call make_from_oop on the final result of the call, but this could be the
2184 // result of merging several call paths. If one of them is made of an actual call node that
2185 // returns an oop, we need to call make_from_oop here as well because we want InlineType
2186 // nodes on every path to avoid merging an unallocated InlineType node path with an oop path.
2187 ret = InlineTypeNode::make_from_oop(this, ret, _gvn.type(ret)->inline_klass());
2188 }
2189 }
2190
2191 return ret;
2192 }
2193
2194 //--------------------set_predefined_input_for_runtime_call--------------------
2195 // Reading and setting the memory state is way conservative here.
2196 // The real problem is that I am not doing real Type analysis on memory,
2197 // so I cannot distinguish card mark stores from other stores. Across a GC
2198 // point the Store Barrier and the card mark memory has to agree. I cannot
2199 // have a card mark store and its barrier split across the GC point from
2200 // either above or below. Here I get that to happen by reading ALL of memory.
2201 // A better answer would be to separate out card marks from other memory.
2202 // For now, return the input memory state, so that it can be reused
2203 // after the call, if this call has restricted memory effects.
2204 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2205 // Set fixed predefined input arguments
2206 call->init_req(TypeFunc::Control, control());
2207 call->init_req(TypeFunc::I_O, top()); // does no i/o
2208 call->init_req(TypeFunc::ReturnAdr, top());
2209 if (call->is_CallLeafPure()) {
2210 call->init_req(TypeFunc::Memory, top());
2272 if (use->is_MergeMem()) {
2273 wl.push(use);
2274 }
2275 }
2276 }
2277
2278 // Replace the call with the current state of the kit.
2279 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2280 JVMState* ejvms = nullptr;
2281 if (has_exceptions()) {
2282 ejvms = transfer_exceptions_into_jvms();
2283 }
2284
2285 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2286 ReplacedNodes replaced_nodes_exception;
2287 Node* ex_ctl = top();
2288
2289 SafePointNode* final_state = stop();
2290
2291 // Find all the needed outputs of this call
2292 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2293
2294 Unique_Node_List wl;
2295 Node* final_mem = final_state->in(TypeFunc::Memory);
2296 Node* final_ctl = final_state->in(TypeFunc::Control);
2297 Node* final_io = final_state->in(TypeFunc::I_O);
2298
2299 // Replace all the old call edges with the edges from the inlining result
2300 if (callprojs->fallthrough_catchproj != nullptr) {
2301 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2302 }
2303 if (callprojs->fallthrough_memproj != nullptr) {
2304 if (final_mem->is_MergeMem()) {
2305 // Parser's exits MergeMem was not transformed but may be optimized
2306 final_mem = _gvn.transform(final_mem);
2307 }
2308 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2309 add_mergemem_users_to_worklist(wl, final_mem);
2310 }
2311 if (callprojs->fallthrough_ioproj != nullptr) {
2312 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2313 }
2314
2315 // Replace the result with the new result if it exists and is used
2316 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2317 // If the inlined code is dead, the result projections for an inline type returned as
2318 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2319 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2320 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2321 "unexpected number of results");
2322 // If we are doing strength reduction and the return type is not loaded we
2323 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2324 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2325 CallNode* new_call = result->in(0)->as_Call();
2326 assert(new_call->proj_out_or_null(TypeFunc::Parms) == result, "the first data projection should be result");
2327 for (uint i = 0; i < callprojs->nb_resproj; i++) {
2328 if (callprojs->resproj[i] != nullptr) {
2329 Node* new_proj = new_call->proj_out_or_null(TypeFunc::Parms + i);
2330 assert(new_proj != nullptr, "projection should be available");
2331 const Type* result_type = _gvn.type(callprojs->resproj[i]);
2332 new_proj = _gvn.transform(new OpaqueParseNode(C, new_proj, result_type));
2333 C->gvn_replace_by(callprojs->resproj[i], new_proj);
2334 }
2335 }
2336 } else {
2337 C->gvn_replace_by(callprojs->resproj[0], result);
2338 }
2339 }
2340
2341 if (ejvms == nullptr) {
2342 // No exception edges to simply kill off those paths
2343 if (callprojs->catchall_catchproj != nullptr) {
2344 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2345 }
2346 if (callprojs->catchall_memproj != nullptr) {
2347 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2348 }
2349 if (callprojs->catchall_ioproj != nullptr) {
2350 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2351 }
2352 // Replace the old exception object with top
2353 if (callprojs->exobj != nullptr) {
2354 C->gvn_replace_by(callprojs->exobj, C->top());
2355 }
2356 } else {
2357 GraphKit ekit(ejvms);
2358
2359 // Load my combined exception state into the kit, with all phis transformed:
2360 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2361 replaced_nodes_exception = ex_map->replaced_nodes();
2362
2363 Node* ex_oop = ekit.use_exception_state(ex_map);
2364
2365 if (callprojs->catchall_catchproj != nullptr) {
2366 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2367 ex_ctl = ekit.control();
2368 }
2369 if (callprojs->catchall_memproj != nullptr) {
2370 Node* ex_mem = ekit.reset_memory();
2371 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2372 add_mergemem_users_to_worklist(wl, ex_mem);
2373 }
2374 if (callprojs->catchall_ioproj != nullptr) {
2375 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2376 }
2377
2378 // Replace the old exception object with the newly created one
2379 if (callprojs->exobj != nullptr) {
2380 C->gvn_replace_by(callprojs->exobj, ex_oop);
2381 }
2382 }
2383
2384 // Disconnect the call from the graph
2385 call->disconnect_inputs(C);
2386 C->gvn_replace_by(call, C->top());
2387
2388 // Clean up any MergeMems that feed other MergeMems since the
2389 // optimizer doesn't like that.
2390 while (wl.size() > 0) {
2391 _gvn.transform(wl.pop());
2392 }
2393
2394 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2395 replaced_nodes.apply(C, final_ctl);
2396 }
2397 if (!ex_ctl->is_top() && do_replaced_nodes) {
2398 replaced_nodes_exception.apply(C, ex_ctl);
2399 }
2400 }
2401
2402
2403 //------------------------------increment_counter------------------------------
2404 // for statistics: increment a VM counter by 1
2405
2406 void GraphKit::increment_counter(address counter_addr) {
2407 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2408 increment_counter(adr1);
2409 }
2410
2411 void GraphKit::increment_counter(Node* counter_addr) {
2412 Node* ctrl = control();
2413 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2414 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2415 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2416 }
2417
2418 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2419 Node* halt = new HaltNode(ctrl, frameptr, reason
2420 PRODUCT_ONLY(COMMA generate_code_in_product));
2421 halt = _gvn.transform(halt);
2422 root()->add_req(halt);
2423 if (_gvn.is_IterGVN() != nullptr) {
2424 record_for_igvn(root());
2425 }
2426 }
2427
2428 //------------------------------uncommon_trap----------------------------------
2429 // Bail out to the interpreter in mid-method. Implemented by calling the
2430 // uncommon_trap blob. This helper function inserts a runtime call with the
2431 // right debug info.
2432 Node* GraphKit::uncommon_trap(int trap_request,
2433 ciKlass* klass, const char* comment,
2434 bool must_throw,
2435 bool keep_exact_action) {
2436 if (failing_internal()) {
2437 stop();
2438 }
2439 if (stopped()) return nullptr; // trap reachable?
2440
2441 // Note: If ProfileTraps is true, and if a deopt. actually
2442 // occurs here, the runtime will make sure an MDO exists. There is
2443 // no need to call method()->ensure_method_data() at this point.
2444
2445 // Set the stack pointer to the right value for reexecution:
2583
2584 /**
2585 * Record profiling data exact_kls for Node n with the type system so
2586 * that it can propagate it (speculation)
2587 *
2588 * @param n node that the type applies to
2589 * @param exact_kls type from profiling
2590 * @param maybe_null did profiling see null?
2591 *
2592 * @return node with improved type
2593 */
2594 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2595 const Type* current_type = _gvn.type(n);
2596 assert(UseTypeSpeculation, "type speculation must be on");
2597
2598 const TypePtr* speculative = current_type->speculative();
2599
2600 // Should the klass from the profile be recorded in the speculative type?
2601 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2602 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2603 const TypeOopPtr* xtype = tklass->as_exact_instance_type();
2604 assert(xtype->klass_is_exact(), "Should be exact");
2605 // Any reason to believe n is not null (from this profiling or a previous one)?
2606 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2607 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2608 // record the new speculative type's depth
2609 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2610 speculative = speculative->with_inline_depth(jvms()->depth());
2611 } else if (current_type->would_improve_ptr(ptr_kind)) {
2612 // Profiling report that null was never seen so we can change the
2613 // speculative type to non null ptr.
2614 if (ptr_kind == ProfileAlwaysNull) {
2615 speculative = TypePtr::NULL_PTR;
2616 } else {
2617 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2618 const TypePtr* ptr = TypePtr::NOTNULL;
2619 if (speculative != nullptr) {
2620 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2621 } else {
2622 speculative = ptr;
2623 }
2624 }
2625 }
2626
2627 if (speculative != current_type->speculative()) {
2628 // Build a type with a speculative type (what we think we know
2629 // about the type but will need a guard when we use it)
2630 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2631 // We're changing the type, we need a new CheckCast node to carry
2632 // the new type. The new type depends on the control: what
2633 // profiling tells us is only valid from here as far as we can
2634 // tell.
2635 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2636 cast = _gvn.transform(cast);
2637 replace_in_map(n, cast);
2638 n = cast;
2639 }
2640
2641 return n;
2642 }
2643
2644 /**
2645 * Record profiling data from receiver profiling at an invoke with the
2646 * type system so that it can propagate it (speculation)
2647 *
2648 * @param n receiver node
2649 *
2650 * @return node with improved type
2651 */
2652 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2653 if (!UseTypeSpeculation) {
2654 return n;
2655 }
2656 ciKlass* exact_kls = profile_has_unique_klass();
2657 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2658 if ((java_bc() == Bytecodes::_checkcast ||
2659 java_bc() == Bytecodes::_instanceof ||
2660 java_bc() == Bytecodes::_aastore) &&
2661 method()->method_data()->is_mature()) {
2662 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2663 if (data != nullptr) {
2664 if (java_bc() == Bytecodes::_aastore) {
2665 ciKlass* array_type = nullptr;
2666 ciKlass* element_type = nullptr;
2667 ProfilePtrKind element_ptr = ProfileMaybeNull;
2668 bool flat_array = true;
2669 bool null_free_array = true;
2670 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2671 exact_kls = element_type;
2672 ptr_kind = element_ptr;
2673 } else {
2674 if (!data->as_BitData()->null_seen()) {
2675 ptr_kind = ProfileNeverNull;
2676 } else {
2677 if (TypeProfileCasts) {
2678 assert(data->is_ReceiverTypeData(), "bad profile data type");
2679 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2680 uint i = 0;
2681 for (; i < call->row_limit(); i++) {
2682 ciKlass* receiver = call->receiver(i);
2683 if (receiver != nullptr) {
2684 break;
2685 }
2686 }
2687 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2688 }
2689 }
2690 }
2691 }
2692 }
2693 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2694 }
2695
2696 /**
2697 * Record profiling data from argument profiling at an invoke with the
2698 * type system so that it can propagate it (speculation)
2699 *
2700 * @param dest_method target method for the call
2701 * @param bc what invoke bytecode is this?
2702 */
2703 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2704 if (!UseTypeSpeculation) {
2705 return;
2706 }
2707 const TypeFunc* tf = TypeFunc::make(dest_method);
2708 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2709 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2710 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2711 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2712 if (is_reference_type(targ->basic_type())) {
2713 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2714 ciKlass* better_type = nullptr;
2715 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2716 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2717 }
2718 i++;
2719 }
2720 }
2721 }
2722
2723 /**
2724 * Record profiling data from parameter profiling at an invoke with
2725 * the type system so that it can propagate it (speculation)
2726 */
2727 void GraphKit::record_profiled_parameters_for_speculation() {
2728 if (!UseTypeSpeculation) {
2729 return;
2730 }
2731 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2851 // The first null ends the list.
2852 Node* parm0, Node* parm1,
2853 Node* parm2, Node* parm3,
2854 Node* parm4, Node* parm5,
2855 Node* parm6, Node* parm7) {
2856 assert(call_addr != nullptr, "must not call null targets");
2857
2858 // Slow-path call
2859 bool is_leaf = !(flags & RC_NO_LEAF);
2860 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2861 if (call_name == nullptr) {
2862 assert(!is_leaf, "must supply name for leaf");
2863 call_name = OptoRuntime::stub_name(call_addr);
2864 }
2865 CallNode* call;
2866 if (!is_leaf) {
2867 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2868 } else if (flags & RC_NO_FP) {
2869 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2870 } else if (flags & RC_VECTOR){
2871 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2872 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2873 } else if (flags & RC_PURE) {
2874 assert(adr_type == nullptr, "pure call does not touch memory");
2875 call = new CallLeafPureNode(call_type, call_addr, call_name);
2876 } else {
2877 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2878 }
2879
2880 // The following is similar to set_edges_for_java_call,
2881 // except that the memory effects of the call are restricted to AliasIdxRaw.
2882
2883 // Slow path call has no side-effects, uses few values
2884 bool wide_in = !(flags & RC_NARROW_MEM);
2885 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2886
2887 Node* prev_mem = nullptr;
2888 if (wide_in) {
2889 prev_mem = set_predefined_input_for_runtime_call(call);
2890 } else {
2891 assert(!wide_out, "narrow in => narrow out");
2892 Node* narrow_mem = memory(adr_type);
2893 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2894 }
2895
2896 // Hook each parm in order. Stop looking at the first null.
2897 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2898 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2899 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2900 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2901 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2902 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2903 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2904 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2905 /* close each nested if ===> */ } } } } } } } }
2906 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2907
2908 if (!is_leaf) {
2909 // Non-leaves can block and take safepoints:
2910 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2911 }
2912 // Non-leaves can throw exceptions:
2913 if (has_io) {
2914 call->set_req(TypeFunc::I_O, i_o());
2915 }
2916
2917 if (flags & RC_UNCOMMON) {
2918 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2919 // (An "if" probability corresponds roughly to an unconditional count.
2920 // Sort of.)
2921 call->set_cnt(PROB_UNLIKELY_MAG(4));
2922 }
2923
2924 Node* c = _gvn.transform(call);
2925 assert(c == call, "cannot disappear");
2926
2934
2935 if (has_io) {
2936 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2937 }
2938 return call;
2939
2940 }
2941
2942 // i2b
2943 Node* GraphKit::sign_extend_byte(Node* in) {
2944 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2945 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2946 }
2947
2948 // i2s
2949 Node* GraphKit::sign_extend_short(Node* in) {
2950 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2951 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2952 }
2953
2954
2955 //------------------------------merge_memory-----------------------------------
2956 // Merge memory from one path into the current memory state.
2957 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2958 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2959 Node* old_slice = mms.force_memory();
2960 Node* new_slice = mms.memory2();
2961 if (old_slice != new_slice) {
2962 PhiNode* phi;
2963 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2964 if (mms.is_empty()) {
2965 // clone base memory Phi's inputs for this memory slice
2966 assert(old_slice == mms.base_memory(), "sanity");
2967 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2968 _gvn.set_type(phi, Type::MEMORY);
2969 for (uint i = 1; i < phi->req(); i++) {
2970 phi->init_req(i, old_slice->in(i));
2971 }
2972 } else {
2973 phi = old_slice->as_Phi(); // Phi was generated already
2974 }
3031 gvn.transform(iff);
3032 if (!bol->is_Con()) gvn.record_for_igvn(iff);
3033 return iff;
3034 }
3035
3036 //-------------------------------gen_subtype_check-----------------------------
3037 // Generate a subtyping check. Takes as input the subtype and supertype.
3038 // Returns 2 values: sets the default control() to the true path and returns
3039 // the false path. Only reads invariant memory; sets no (visible) memory.
3040 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
3041 // but that's not exposed to the optimizer. This call also doesn't take in an
3042 // Object; if you wish to check an Object you need to load the Object's class
3043 // prior to coming here.
3044 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
3045 ciMethod* method, int bci) {
3046 Compile* C = gvn.C;
3047 if ((*ctrl)->is_top()) {
3048 return C->top();
3049 }
3050
3051 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
3052 // For a direct pointer comparison, we need the refined array klass pointer
3053 Node* vm_superklass = superklass;
3054 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
3055 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
3056 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
3057 }
3058
3059 // Fast check for identical types, perhaps identical constants.
3060 // The types can even be identical non-constants, in cases
3061 // involving Array.newInstance, Object.clone, etc.
3062 if (subklass == superklass)
3063 return C->top(); // false path is dead; no test needed.
3064
3065 if (gvn.type(superklass)->singleton()) {
3066 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3067 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
3068
3069 // In the common case of an exact superklass, try to fold up the
3070 // test before generating code. You may ask, why not just generate
3071 // the code and then let it fold up? The answer is that the generated
3072 // code will necessarily include null checks, which do not always
3073 // completely fold away. If they are also needless, then they turn
3074 // into a performance loss. Example:
3075 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3076 // Here, the type of 'fa' is often exact, so the store check
3077 // of fa[1]=x will fold up, without testing the nullness of x.
3078 //
3079 // At macro expansion, we would have already folded the SubTypeCheckNode
3080 // being expanded here because we always perform the static sub type
3081 // check in SubTypeCheckNode::sub() regardless of whether
3082 // StressReflectiveCode is set or not. We can therefore skip this
3083 // static check when StressReflectiveCode is on.
3084 switch (C->static_subtype_check(superk, subk)) {
3085 case Compile::SSC_always_false:
3086 {
3087 Node* always_fail = *ctrl;
3088 *ctrl = gvn.C->top();
3089 return always_fail;
3090 }
3091 case Compile::SSC_always_true:
3092 return C->top();
3093 case Compile::SSC_easy_test:
3094 {
3095 // Just do a direct pointer compare and be done.
3096 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3097 *ctrl = gvn.transform(new IfTrueNode(iff));
3098 return gvn.transform(new IfFalseNode(iff));
3099 }
3100 case Compile::SSC_full_test:
3101 break;
3102 default:
3103 ShouldNotReachHere();
3104 }
3105 }
3106
3107 // %%% Possible further optimization: Even if the superklass is not exact,
3108 // if the subklass is the unique subtype of the superklass, the check
3109 // will always succeed. We could leave a dependency behind to ensure this.
3110
3111 // First load the super-klass's check-offset
3112 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3113 Node* m = C->immutable_memory();
3114 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3115 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3116 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3154 gvn.record_for_igvn(r_ok_subtype);
3155
3156 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3157 // SubTypeCheck node
3158 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3159 ciCallProfile profile = method->call_profile_at_bci(bci);
3160 float total_prob = 0;
3161 for (int i = 0; profile.has_receiver(i); ++i) {
3162 float prob = profile.receiver_prob(i);
3163 total_prob += prob;
3164 }
3165 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3166 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3167 for (int i = 0; profile.has_receiver(i); ++i) {
3168 ciKlass* klass = profile.receiver(i);
3169 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3170 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3171 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3172 continue;
3173 }
3174 if (klass_t->isa_aryklassptr()) {
3175 // For a direct pointer comparison, we need the refined array klass pointer
3176 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3177 }
3178 float prob = profile.receiver_prob(i);
3179 ConNode* klass_node = gvn.makecon(klass_t);
3180 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3181 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3182
3183 if (result == Compile::SSC_always_true) {
3184 r_ok_subtype->add_req(iftrue);
3185 } else {
3186 assert(result == Compile::SSC_always_false, "");
3187 r_not_subtype->add_req(iftrue);
3188 }
3189 *ctrl = gvn.transform(new IfFalseNode(iff));
3190 }
3191 }
3192 }
3193
3194 // See if we get an immediate positive hit. Happens roughly 83% of the
3195 // time. Test to see if the value loaded just previously from the subklass
3196 // is exactly the superklass.
3197 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3211 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3212 }
3213 return not_subtype_ctrl;
3214 }
3215
3216 r_ok_subtype->init_req(1, iftrue1);
3217
3218 // Check for immediate negative hit. Happens roughly 11% of the time (which
3219 // is roughly 63% of the remaining cases). Test to see if the loaded
3220 // check-offset points into the subklass display list or the 1-element
3221 // cache. If it points to the display (and NOT the cache) and the display
3222 // missed then it's not a subtype.
3223 Node *cacheoff = gvn.intcon(cacheoff_con);
3224 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3225 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3226 *ctrl = gvn.transform(new IfFalseNode(iff2));
3227
3228 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3229 // No performance impact (too rare) but allows sharing of secondary arrays
3230 // which has some footprint reduction.
3231 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3232 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3233 *ctrl = gvn.transform(new IfFalseNode(iff3));
3234
3235 // -- Roads not taken here: --
3236 // We could also have chosen to perform the self-check at the beginning
3237 // of this code sequence, as the assembler does. This would not pay off
3238 // the same way, since the optimizer, unlike the assembler, can perform
3239 // static type analysis to fold away many successful self-checks.
3240 // Non-foldable self checks work better here in second position, because
3241 // the initial primary superclass check subsumes a self-check for most
3242 // types. An exception would be a secondary type like array-of-interface,
3243 // which does not appear in its own primary supertype display.
3244 // Finally, we could have chosen to move the self-check into the
3245 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3246 // dependent manner. But it is worthwhile to have the check here,
3247 // where it can be perhaps be optimized. The cost in code space is
3248 // small (register compare, branch).
3249
3250 // Now do a linear scan of the secondary super-klass array. Again, no real
3251 // performance impact (too rare) but it's gotta be done.
3252 // Since the code is rarely used, there is no penalty for moving it
3253 // out of line, and it can only improve I-cache density.
3254 // The decision to inline or out-of-line this final check is platform
3255 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3256 Node* psc = gvn.transform(
3257 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3258
3259 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3260 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3261 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3262
3263 // Return false path; set default control to true path.
3264 *ctrl = gvn.transform(r_ok_subtype);
3265 return gvn.transform(r_not_subtype);
3266 }
3267
3268 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3269 const Type* sub_t = _gvn.type(obj_or_subklass);
3270 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3271 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3272 obj_or_subklass = makecon(sub_t);
3273 }
3274 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3275 if (expand_subtype_check) {
3276 MergeMemNode* mem = merged_memory();
3277 Node* ctrl = control();
3278 Node* subklass = obj_or_subklass;
3279 if (!sub_t->isa_klassptr()) {
3280 subklass = load_object_klass(obj_or_subklass);
3281 }
3282
3283 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3284 set_control(ctrl);
3285 return n;
3286 }
3287
3288 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3289 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3290 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3291 set_control(_gvn.transform(new IfTrueNode(iff)));
3292 return _gvn.transform(new IfFalseNode(iff));
3293 }
3294
3295 // Profile-driven exact type check:
3296 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3297 float prob, Node* *casted_receiver) {
3298 assert(!klass->is_interface(), "no exact type check on interfaces");
3299 Node* fail = top();
3300 const Type* rec_t = _gvn.type(receiver);
3301 if (rec_t->is_inlinetypeptr()) {
3302 if (klass->equals(rec_t->inline_klass())) {
3303 (*casted_receiver) = receiver; // Always passes
3304 } else {
3305 (*casted_receiver) = top(); // Always fails
3306 fail = control();
3307 set_control(top());
3308 }
3309 return fail;
3310 }
3311 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3312 if (tklass->isa_aryklassptr()) {
3313 // For a direct pointer comparison, we need the refined array klass pointer
3314 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3315 }
3316 Node* recv_klass = load_object_klass(receiver);
3317 fail = type_check(recv_klass, tklass, prob);
3318
3319 if (!stopped()) {
3320 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3321 const TypeOopPtr* recv_xtype = tklass->as_exact_instance_type();
3322 assert(recv_xtype->klass_is_exact(), "");
3323
3324 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3325 // Subsume downstream occurrences of receiver with a cast to
3326 // recv_xtype, since now we know what the type will be.
3327 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3328 Node* res = _gvn.transform(cast);
3329 if (recv_xtype->is_inlinetypeptr()) {
3330 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3331 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3332 }
3333 (*casted_receiver) = res;
3334 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3335 // (User must make the replace_in_map call.)
3336 }
3337 }
3338
3339 return fail;
3340 }
3341
3342 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3343 float prob) {
3344 Node* want_klass = makecon(tklass);
3345 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3346 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3347 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3348 set_control(_gvn.transform(new IfTrueNode (iff)));
3349 Node* fail = _gvn.transform(new IfFalseNode(iff));
3350 return fail;
3351 }
3352
3353 //------------------------------subtype_check_receiver-------------------------
3354 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3355 Node** casted_receiver) {
3356 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3357 Node* want_klass = makecon(tklass);
3358
3359 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3360
3361 // Ignore interface type information until interface types are properly tracked.
3362 if (!stopped() && !klass->is_interface()) {
3363 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3364 const TypeOopPtr* recv_type = tklass->as_subtype_instance_type();
3365 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3366 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3367 if (recv_type->is_inlinetypeptr()) {
3368 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3369 }
3370 (*casted_receiver) = cast;
3371 }
3372 }
3373
3374 return slow_ctl;
3375 }
3376
3377 //------------------------------seems_never_null-------------------------------
3378 // Use null_seen information if it is available from the profile.
3379 // If we see an unexpected null at a type check we record it and force a
3380 // recompile; the offending check will be recompiled to handle nulls.
3381 // If we see several offending BCIs, then all checks in the
3382 // method will be recompiled.
3383 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3384 speculating = !_gvn.type(obj)->speculative_maybe_null();
3385 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3386 if (UncommonNullCast // Cutout for this technique
3387 && obj != null() // And not the -Xcomp stupid case?
3388 && !too_many_traps(reason)
3389 ) {
3390 if (speculating) {
3459
3460 //------------------------maybe_cast_profiled_receiver-------------------------
3461 // If the profile has seen exactly one type, narrow to exactly that type.
3462 // Subsequent type checks will always fold up.
3463 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3464 const TypeKlassPtr* require_klass,
3465 ciKlass* spec_klass,
3466 bool safe_for_replace) {
3467 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3468
3469 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3470
3471 // Make sure we haven't already deoptimized from this tactic.
3472 if (too_many_traps_or_recompiles(reason))
3473 return nullptr;
3474
3475 // (No, this isn't a call, but it's enough like a virtual call
3476 // to use the same ciMethod accessor to get the profile info...)
3477 // If we have a speculative type use it instead of profiling (which
3478 // may not help us)
3479 ciKlass* exact_kls = spec_klass;
3480 if (exact_kls == nullptr) {
3481 if (java_bc() == Bytecodes::_aastore) {
3482 ciKlass* array_type = nullptr;
3483 ciKlass* element_type = nullptr;
3484 ProfilePtrKind element_ptr = ProfileMaybeNull;
3485 bool flat_array = true;
3486 bool null_free_array = true;
3487 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3488 exact_kls = element_type;
3489 } else {
3490 exact_kls = profile_has_unique_klass();
3491 }
3492 }
3493 if (exact_kls != nullptr) {// no cast failures here
3494 if (require_klass == nullptr ||
3495 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3496 // If we narrow the type to match what the type profile sees or
3497 // the speculative type, we can then remove the rest of the
3498 // cast.
3499 // This is a win, even if the exact_kls is very specific,
3500 // because downstream operations, such as method calls,
3501 // will often benefit from the sharper type.
3502 Node* exact_obj = not_null_obj; // will get updated in place...
3503 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3504 &exact_obj);
3505 { PreserveJVMState pjvms(this);
3506 set_control(slow_ctl);
3507 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3508 }
3509 if (safe_for_replace) {
3510 replace_in_map(not_null_obj, exact_obj);
3511 }
3512 return exact_obj;
3602
3603 // If not_null_obj is dead, only null-path is taken
3604 if (stopped()) { // Doing instance-of on a null?
3605 set_control(null_ctl);
3606 return intcon(0);
3607 }
3608 region->init_req(_null_path, null_ctl);
3609 phi ->init_req(_null_path, intcon(0)); // Set null path value
3610 if (null_ctl == top()) {
3611 // Do this eagerly, so that pattern matches like is_diamond_phi
3612 // will work even during parsing.
3613 assert(_null_path == PATH_LIMIT-1, "delete last");
3614 region->del_req(_null_path);
3615 phi ->del_req(_null_path);
3616 }
3617
3618 // Do we know the type check always succeed?
3619 bool known_statically = false;
3620 if (improved_klass_ptr_type->singleton()) {
3621 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3622 if (subk != nullptr && subk->is_loaded()) {
3623 int static_res = C->static_subtype_check(improved_klass_ptr_type, subk);
3624 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3625 }
3626 }
3627
3628 if (!known_statically) {
3629 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3630 // We may not have profiling here or it may not help us. If we
3631 // have a speculative type use it to perform an exact cast.
3632 ciKlass* spec_obj_type = obj_type->speculative_type();
3633 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3634 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3635 if (stopped()) { // Profile disagrees with this path.
3636 set_control(null_ctl); // Null is the only remaining possibility.
3637 return intcon(0);
3638 }
3639 if (cast_obj != nullptr) {
3640 not_null_obj = cast_obj;
3641 }
3642 }
3662 record_for_igvn(region);
3663
3664 // If we know the type check always succeeds then we don't use the
3665 // profiling data at this bytecode. Don't lose it, feed it to the
3666 // type system as a speculative type.
3667 if (safe_for_replace) {
3668 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3669 replace_in_map(obj, casted_obj);
3670 }
3671
3672 return _gvn.transform(phi);
3673 }
3674
3675 //-------------------------------gen_checkcast---------------------------------
3676 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3677 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3678 // uncommon-trap paths work. Adjust stack after this call.
3679 // If failure_control is supplied and not null, it is filled in with
3680 // the control edge for the cast failure. Otherwise, an appropriate
3681 // uncommon trap or exception is thrown.
3682 // If 'new_cast_failure_map' is supplied and is not null, it is set to a newly cloned map
3683 // when the current map for the success path is updated with information only present
3684 // on the success path and not the cast failure path. The newly cloned map should then be
3685 // used to emit the uncommon trap in the caller.
3686 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node** failure_control, SafePointNode** new_cast_failure_map, bool null_free, bool maybe_larval) {
3687 assert(new_cast_failure_map == nullptr || failure_control != nullptr,
3688 "failure_control must be set when new_failure_map is used");
3689 kill_dead_locals(); // Benefit all the uncommon traps
3690 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3691 const Type* obj_type = _gvn.type(obj);
3692
3693 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3694 const TypeOopPtr* toop = improved_klass_ptr_type->as_subtype_instance_type();
3695 bool safe_for_replace = (failure_control == nullptr);
3696 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3697
3698 // Fast cutout: Check the case that the cast is vacuously true.
3699 // This detects the common cases where the test will short-circuit
3700 // away completely. We do this before we perform the null check,
3701 // because if the test is going to turn into zero code, we don't
3702 // want a residual null check left around. (Causes a slowdown,
3703 // for example, in some objArray manipulations, such as a[i]=a[j].)
3704 if (improved_klass_ptr_type->singleton()) {
3705 const TypeKlassPtr* kptr = nullptr;
3706 if (obj_type->isa_oop_ptr()) {
3707 kptr = obj_type->is_oopptr()->as_klass_type();
3708 } else if (obj->is_InlineType()) {
3709 ciInlineKlass* vk = obj_type->inline_klass();
3710 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3711 }
3712
3713 if (kptr != nullptr) {
3714 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3715 case Compile::SSC_always_true:
3716 // If we know the type check always succeed then we don't use
3717 // the profiling data at this bytecode. Don't lose it, feed it
3718 // to the type system as a speculative type.
3719 obj = record_profiled_receiver_for_speculation(obj);
3720 if (null_free) {
3721 assert(safe_for_replace, "must be");
3722 obj = null_check(obj);
3723 }
3724 return obj;
3725 case Compile::SSC_always_false:
3726 if (null_free) {
3727 assert(safe_for_replace, "must be");
3728 obj = null_check(obj);
3729 }
3730 // It needs a null check because a null will *pass* the cast check.
3731 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3732 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3733 Deoptimization::DeoptReason reason = is_aastore ?
3734 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3735 builtin_throw(reason);
3736 return top();
3737 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3738 return null_assert(obj);
3739 }
3740 break; // Fall through to full check
3741 default:
3742 break;
3743 }
3744 }
3745 }
3746
3747 ciProfileData* data = nullptr;
3748 if (failure_control == nullptr) { // use MDO in regular case only
3749 assert(java_bc() == Bytecodes::_aastore ||
3750 java_bc() == Bytecodes::_checkcast,
3751 "interpreter profiles type checks only for these BCs");
3752 if (method()->method_data()->is_mature()) {
3753 data = method()->method_data()->bci_to_data(bci());
3754 }
3755 }
3756
3757 // Make the merge point
3758 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3759 RegionNode* region = new RegionNode(PATH_LIMIT);
3760 Node* phi = new PhiNode(region, toop);
3761 _gvn.set_type(region, Type::CONTROL);
3762 _gvn.set_type(phi, toop);
3763
3764 C->set_has_split_ifs(true); // Has chance for split-if optimization
3765
3766 // Use null-cast information if it is available
3767 bool speculative_not_null = false;
3768 bool never_see_null = ((failure_control == nullptr) // regular case only
3769 && seems_never_null(obj, data, speculative_not_null));
3770
3771 // Null check; get casted pointer; set region slot 3
3772 Node* null_ctl = top();
3773 Node* not_null_obj = nullptr;
3774 if (null_free) {
3775 assert(safe_for_replace, "must be");
3776 not_null_obj = null_check(obj);
3777 } else {
3778 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, false /*safe_for_replace*/, speculative_not_null);
3779 }
3780
3781 // If not_null_obj is dead, only null-path is taken
3782 if (stopped()) { // Doing instance-of on a null?
3783 set_control(null_ctl);
3784 if (toop->is_inlinetypeptr()) {
3785 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3786 }
3787 return null();
3788 }
3789 region->init_req(_null_path, null_ctl);
3790 phi ->init_req(_null_path, null()); // Set null path value
3791 if (null_ctl == top()) {
3792 // Do this eagerly, so that pattern matches like is_diamond_phi
3793 // will work even during parsing.
3794 assert(_null_path == PATH_LIMIT-1, "delete last");
3795 region->del_req(_null_path);
3796 phi ->del_req(_null_path);
3797 }
3798
3799 Node* cast_obj = nullptr;
3800 if (improved_klass_ptr_type->klass_is_exact()) {
3801 // The following optimization tries to statically cast the speculative type of the object
3802 // (for example obtained during profiling) to the type of the superklass and then do a
3803 // dynamic check that the type of the object is what we expect. To work correctly
3804 // for checkcast and aastore the type of superklass should be exact.
3805 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3806 // We may not have profiling here or it may not help us. If we have
3807 // a speculative type use it to perform an exact cast.
3808 ciKlass* spec_obj_type = obj_type->speculative_type();
3809 if (spec_obj_type != nullptr || data != nullptr) {
3810 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, false /*safe_for_replace*/);
3811 if (cast_obj != nullptr) {
3812 if (failure_control != nullptr) // failure is now impossible
3813 (*failure_control) = top();
3814 // adjust the type of the phi to the exact klass:
3815 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3816 }
3817 }
3818 }
3819
3820 if (cast_obj == nullptr) {
3821 // Generate the subtype check
3822 Node* improved_superklass = superklass;
3823 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3824 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3825 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3826 // Additionally, the benefit would only be minor in non-constant cases.
3827 improved_superklass = makecon(improved_klass_ptr_type);
3828 }
3829 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3830 // Plug in success path into the merge
3831 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3832 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3833 if (failure_control == nullptr) {
3834 if (not_subtype_ctrl != top()) { // If failure is possible
3835 PreserveJVMState pjvms(this);
3836 set_control(not_subtype_ctrl);
3837 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3838 Deoptimization::DeoptReason reason = is_aastore ?
3839 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3840 builtin_throw(reason);
3841 }
3842 } else {
3843 (*failure_control) = not_subtype_ctrl;
3844 }
3845 }
3846
3847 region->init_req(_obj_path, control());
3848 phi ->init_req(_obj_path, cast_obj);
3849
3850 // Return final merged results
3851 set_control( _gvn.transform(region) );
3852 record_for_igvn(region);
3853
3854 bool not_inline = !toop->can_be_inline_type();
3855 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3856 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3857 // Check if obj has been loaded from an array. Keep obj unchanged for final
3858 // map replacement below.
3859 Node* array_obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3860 Node* array = nullptr;
3861 if (array_obj->isa_Load()) {
3862 Node* address = array_obj->in(MemNode::Address);
3863 if (address->isa_AddP()) {
3864 array = address->as_AddP()->in(AddPNode::Base);
3865 }
3866 } else if (array_obj->is_Phi()) {
3867 Node* region = array_obj->in(0);
3868 // TODO make this more robust (see JDK-8231346)
3869 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3870 IfNode* iff = region->in(2)->in(0)->isa_If();
3871 if (iff != nullptr) {
3872 iff->is_flat_array_check(&_gvn, &array);
3873 }
3874 }
3875 }
3876 if (array != nullptr) {
3877 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3878 if (ary_t != nullptr) {
3879 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3880 // Casting array element to a non-inline-type, mark array as not null-free.
3881 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3882 if (new_cast_failure_map != nullptr) {
3883 // We want to propagate the improved cast node in the current map. Clone it such that we can still properly
3884 // create the cast failure path in the caller without wrongly making the cast node live there.
3885 *new_cast_failure_map = clone_map();
3886 }
3887 replace_in_map(array, cast);
3888 array = cast;
3889 }
3890 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3891 // Casting array element to a non-flat-in-array type, mark array as not flat.
3892 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3893 if (new_cast_failure_map != nullptr && *new_cast_failure_map == nullptr) {
3894 // Same as above.
3895 *new_cast_failure_map = clone_map();
3896 }
3897 replace_in_map(array, cast);
3898 array = cast;
3899 }
3900 }
3901 }
3902 }
3903
3904 // A merge of null or Casted-NotNull obj
3905 Node* res = _gvn.transform(phi);
3906 if (!stopped() && !res->is_InlineType()) {
3907 res = record_profiled_receiver_for_speculation(res);
3908 if (toop->is_inlinetypeptr() && !maybe_larval) {
3909 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3910 res = vt;
3911 if (safe_for_replace) {
3912 replace_in_map(obj, vt);
3913 replace_in_map(not_null_obj, vt);
3914 replace_in_map(res, vt);
3915 }
3916 } else if (safe_for_replace) {
3917 replace_in_map(obj, res);
3918 }
3919 }
3920 return res;
3921 }
3922
3923 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3924 // Load markword
3925 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3926 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3927 if (check_lock && !UseCompactObjectHeaders) {
3928 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3929 // Check if obj is locked
3930 Node* locked_bit = MakeConX(markWord::unlocked_value);
3931 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3932 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3933 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3934 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3935 _gvn.transform(iff);
3936 Node* locked_region = new RegionNode(3);
3937 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3938
3939 // Unlocked: Use bits from mark word
3940 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3941 mark_phi->init_req(1, mark);
3942
3943 // Locked: Load prototype header from klass
3944 set_control(_gvn.transform(new IfFalseNode(iff)));
3945 // Make loads control dependent to make sure they are only executed if array is locked
3946 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3947 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3948 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3949 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3950
3951 locked_region->init_req(2, control());
3952 mark_phi->init_req(2, proto);
3953 set_control(_gvn.transform(locked_region));
3954 record_for_igvn(locked_region);
3955
3956 mark = mark_phi;
3957 }
3958
3959 // Now check if mark word bits are set
3960 Node* mask = MakeConX(mask_val);
3961 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3962 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3963 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3964 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3965 }
3966
3967 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3968 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3969 }
3970
3971 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3972 // We can't use immutable memory here because the mark word is mutable.
3973 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3974 // check is moved out of loops (mainly to enable loop unswitching).
3975 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3976 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3977 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3978 }
3979
3980 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3981 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3982 }
3983
3984 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3985 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3986
3987 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3988 if (!vk->has_null_free_non_atomic_layout()) {
3989 return intcon(1); // Always atomic
3990 } else if (!vk->has_null_free_atomic_layout()) {
3991 return intcon(0); // Never atomic
3992 }
3993
3994 Node* array_klass = load_object_klass(array);
3995 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3996 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
3997 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3998 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3999 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4000 }
4001
4002 Node* GraphKit::atomic_layout_array_test_and_get_layout_kind(Node* array, RegionNode* atomic_region) {
4003 Node* array_klass = load_object_klass(array);
4004 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
4005 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
4006 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
4007 Node* cmp_null_free = _gvn.transform(new CmpINode(layout_kind, intcon(static_cast<jint>(LayoutKind::NULL_FREE_ATOMIC_FLAT))));
4008 Node* bol_null_free = _gvn.transform(new BoolNode(cmp_null_free, BoolTest::eq));
4009 Node* cmp_nullable = _gvn.transform(new CmpINode(layout_kind, intcon(static_cast<jint>(LayoutKind::NULLABLE_ATOMIC_FLAT))));
4010 Node* bol_nullable = _gvn.transform(new BoolNode(cmp_nullable, BoolTest::eq));
4011
4012 IfNode* iff_null_free = create_and_xform_if(control(), bol_null_free, PROB_FAIR, COUNT_UNKNOWN);
4013 atomic_region->add_req(_gvn.transform(new IfTrueNode(iff_null_free)));
4014 set_control(_gvn.transform(new IfFalseNode(iff_null_free)));
4015
4016 IfNode* iff_nullable = create_and_xform_if(control(), bol_nullable, PROB_FAIR, COUNT_UNKNOWN);
4017 atomic_region->add_req(_gvn.transform(new IfTrueNode(iff_nullable)));
4018 set_control(_gvn.transform(new IfFalseNode(iff_nullable)));
4019
4020 return layout_kind;
4021 }
4022
4023 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
4024 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
4025 RegionNode* region = new RegionNode(3);
4026 Node* null_ctl = top();
4027 null_check_oop(val, &null_ctl);
4028 if (null_ctl != top()) {
4029 PreserveJVMState pjvms(this);
4030 set_control(null_ctl);
4031 {
4032 // Deoptimize if null-free array
4033 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
4034 inc_sp(nargs);
4035 uncommon_trap(Deoptimization::Reason_null_check,
4036 Deoptimization::Action_none);
4037 }
4038 region->init_req(1, control());
4039 }
4040 region->init_req(2, control());
4041 set_control(_gvn.transform(region));
4042 record_for_igvn(region);
4043 if (_gvn.type(val) == TypePtr::NULL_PTR) {
4044 // Since we were just successfully storing null, the array can't be null free.
4045 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
4046 ary_t = ary_t->cast_to_not_null_free();
4047 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
4048 if (safe_for_replace) {
4049 replace_in_map(ary, cast);
4050 }
4051 ary = cast;
4052 }
4053 return ary;
4054 }
4055
4056 //------------------------------next_monitor-----------------------------------
4057 // What number should be given to the next monitor?
4058 int GraphKit::next_monitor() {
4059 int current = jvms()->monitor_depth()* C->sync_stack_slots();
4060 int next = current + C->sync_stack_slots();
4061 // Keep the toplevel high water mark current:
4062 if (C->fixed_slots() < next) C->set_fixed_slots(next);
4063 return current;
4064 }
4065
4066 //------------------------------insert_mem_bar---------------------------------
4067 // Memory barrier to avoid floating things around
4068 // The membar serves as a pinch point between both control and all memory slices.
4069 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
4070 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
4071 mb->init_req(TypeFunc::Control, control());
4072 mb->init_req(TypeFunc::Memory, reset_memory());
4073 Node* membar = _gvn.transform(mb);
4074 record_for_igvn(membar);
4075 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4176 lock->create_lock_counter(map()->jvms());
4177 increment_counter(lock->counter()->addr());
4178 }
4179 #endif
4180
4181 return flock;
4182 }
4183
4184
4185 //------------------------------shared_unlock----------------------------------
4186 // Emit unlocking code.
4187 void GraphKit::shared_unlock(Node* box, Node* obj) {
4188 // bci is either a monitorenter bc or InvocationEntryBci
4189 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4190 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4191
4192 if (stopped()) { // Dead monitor?
4193 map()->pop_monitor(); // Kill monitor from debug info
4194 return;
4195 }
4196 assert(!obj->is_InlineType(), "should not unlock on inline type");
4197
4198 // Memory barrier to avoid floating things down past the locked region
4199 insert_mem_bar(Op_MemBarReleaseLock);
4200
4201 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4202 UnlockNode *unlock = new UnlockNode(C, tf);
4203 #ifdef ASSERT
4204 unlock->set_dbg_jvms(sync_jvms());
4205 #endif
4206 uint raw_idx = Compile::AliasIdxRaw;
4207 unlock->init_req( TypeFunc::Control, control() );
4208 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4209 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4210 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4211 unlock->init_req( TypeFunc::ReturnAdr, top() );
4212
4213 unlock->init_req(TypeFunc::Parms + 0, obj);
4214 unlock->init_req(TypeFunc::Parms + 1, box);
4215 unlock = _gvn.transform(unlock)->as_Unlock();
4216
4217 Node* mem = reset_memory();
4218
4219 // unlock has no side-effects, sets few values
4220 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4221
4222 // Kill monitor from debug info
4223 map()->pop_monitor( );
4224 }
4225
4226 //-------------------------------get_layout_helper-----------------------------
4227 // If the given klass is a constant or known to be an array,
4228 // fetch the constant layout helper value into constant_value
4229 // and return null. Otherwise, load the non-constant
4230 // layout helper value, and return the node which represents it.
4231 // This two-faced routine is useful because allocation sites
4232 // almost always feature constant types.
4233 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4234 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4235 if (!StressReflectiveCode && klass_t != nullptr) {
4236 bool xklass = klass_t->klass_is_exact();
4237 bool can_be_flat = false;
4238 const TypeAryPtr* ary_type = klass_t->as_exact_instance_type()->isa_aryptr();
4239 if (UseArrayFlattening && !xklass && ary_type != nullptr) {
4240 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4241 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4242 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4243 }
4244 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4245 jint lhelper;
4246 if (klass_t->is_flat()) {
4247 lhelper = ary_type->flat_layout_helper();
4248 } else if (klass_t->isa_aryklassptr()) {
4249 BasicType elem = ary_type->elem()->array_element_basic_type();
4250 if (is_reference_type(elem, true)) {
4251 elem = T_OBJECT;
4252 }
4253 lhelper = Klass::array_layout_helper(elem);
4254 } else {
4255 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4256 }
4257 if (lhelper != Klass::_lh_neutral_value) {
4258 constant_value = lhelper;
4259 return (Node*) nullptr;
4260 }
4261 }
4262 }
4263 constant_value = Klass::_lh_neutral_value; // put in a known value
4264 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
4265 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4266 }
4267
4268 // We just put in an allocate/initialize with a big raw-memory effect.
4269 // Hook selected additional alias categories on the initialization.
4270 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4271 MergeMemNode* init_in_merge,
4272 Node* init_out_raw) {
4273 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4274 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4275
4276 Node* prevmem = kit.memory(alias_idx);
4277 init_in_merge->set_memory_at(alias_idx, prevmem);
4278 if (init_out_raw != nullptr) {
4279 kit.set_memory(init_out_raw, alias_idx);
4280 }
4281 }
4282
4283 //---------------------------set_output_for_allocation-------------------------
4284 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4285 const TypeOopPtr* oop_type,
4286 bool deoptimize_on_exception) {
4287 int rawidx = Compile::AliasIdxRaw;
4288 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4289 add_safepoint_edges(alloc);
4290 Node* allocx = _gvn.transform(alloc);
4291 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4292 // create memory projection for i_o
4293 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4294 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4295
4296 // create a memory projection as for the normal control path
4297 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4298 set_memory(malloc, rawidx);
4299
4300 // a normal slow-call doesn't change i_o, but an allocation does
4301 // we create a separate i_o projection for the normal control path
4302 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4303 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4304
4305 // put in an initialization barrier
4306 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4307 rawoop)->as_Initialize();
4308 assert(alloc->initialization() == init, "2-way macro link must work");
4309 assert(init ->allocation() == alloc, "2-way macro link must work");
4310 {
4311 // Extract memory strands which may participate in the new object's
4312 // initialization, and source them from the new InitializeNode.
4313 // This will allow us to observe initializations when they occur,
4314 // and link them properly (as a group) to the InitializeNode.
4315 assert(init->in(InitializeNode::Memory) == malloc, "");
4316 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4317 init->set_req(InitializeNode::Memory, minit_in);
4318 record_for_igvn(minit_in); // fold it up later, if possible
4319 _gvn.set_type(minit_in, Type::MEMORY);
4320 Node* minit_out = memory(rawidx);
4321 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4322 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4323 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4324 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4325 // multiple projections as a result.
4326 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4327 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4328 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4329 if (oop_type->isa_aryptr()) {
4330 // Initially all flat array accesses share a single slice
4331 // but that changes after parsing. Prepare the memory graph so
4332 // it can optimize flat array accesses properly once they
4333 // don't share a single slice.
4334 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4335 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4336 int elemidx = C->get_alias_index(telemref);
4337 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4338 if (alias_adr_type->is_flat()) {
4339 C->set_flat_accesses();
4340 }
4341 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4342 } else if (oop_type->isa_instptr()) {
4343 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4344 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4345 ciField* field = ik->nonstatic_field_at(i);
4346 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4347 continue; // do not bother to track really large numbers of fields
4348 // Find (or create) the alias category for this field:
4349 int fieldidx = C->alias_type(field)->index();
4350 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4351 }
4352 }
4353 }
4354
4355 // Cast raw oop to the real thing...
4356 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4357 javaoop = _gvn.transform(javaoop);
4358 C->set_recent_alloc(control(), javaoop);
4359 assert(just_allocated_object(control()) == javaoop, "just allocated");
4360
4361 #ifdef ASSERT
4373 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4374 }
4375 }
4376 #endif //ASSERT
4377
4378 return javaoop;
4379 }
4380
4381 //---------------------------new_instance--------------------------------------
4382 // This routine takes a klass_node which may be constant (for a static type)
4383 // or may be non-constant (for reflective code). It will work equally well
4384 // for either, and the graph will fold nicely if the optimizer later reduces
4385 // the type to a constant.
4386 // The optional arguments are for specialized use by intrinsics:
4387 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4388 // - If 'return_size_val', report the total object size to the caller.
4389 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4390 Node* GraphKit::new_instance(Node* klass_node,
4391 Node* extra_slow_test,
4392 Node* *return_size_val,
4393 bool deoptimize_on_exception,
4394 InlineTypeNode* inline_type_node) {
4395 // Compute size in doublewords
4396 // The size is always an integral number of doublewords, represented
4397 // as a positive bytewise size stored in the klass's layout_helper.
4398 // The layout_helper also encodes (in a low bit) the need for a slow path.
4399 jint layout_con = Klass::_lh_neutral_value;
4400 Node* layout_val = get_layout_helper(klass_node, layout_con);
4401 bool layout_is_con = (layout_val == nullptr);
4402
4403 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4404 // Generate the initial go-slow test. It's either ALWAYS (return a
4405 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4406 // case) a computed value derived from the layout_helper.
4407 Node* initial_slow_test = nullptr;
4408 if (layout_is_con) {
4409 assert(!StressReflectiveCode, "stress mode does not use these paths");
4410 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4411 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4412 } else { // reflective case
4413 // This reflective path is used by Unsafe.allocateInstance.
4414 // (It may be stress-tested by specifying StressReflectiveCode.)
4415 // Basically, we want to get into the VM is there's an illegal argument.
4416 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4417 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4418 if (extra_slow_test != intcon(0)) {
4419 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4420 }
4421 // (Macro-expander will further convert this to a Bool, if necessary.)
4427 if (layout_is_con) {
4428 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4429 } else { // reflective case
4430 // This reflective path is used by clone and Unsafe.allocateInstance.
4431 size = ConvI2X(layout_val);
4432
4433 // Clear the low bits to extract layout_helper_size_in_bytes:
4434 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4435 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4436 size = _gvn.transform( new AndXNode(size, mask) );
4437 }
4438 if (return_size_val != nullptr) {
4439 (*return_size_val) = size;
4440 }
4441
4442 // This is a precise notnull oop of the klass.
4443 // (Actually, it need not be precise if this is a reflective allocation.)
4444 // It's what we cast the result to.
4445 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4446 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4447 const TypeOopPtr* oop_type = tklass->as_exact_instance_type();
4448
4449 // Now generate allocation code
4450
4451 // The entire memory state is needed for slow path of the allocation
4452 // since GC and deoptimization can happen.
4453 Node *mem = reset_memory();
4454 set_all_memory(mem); // Create new memory state
4455
4456 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4457 control(), mem, i_o(),
4458 size, klass_node,
4459 initial_slow_test, inline_type_node);
4460
4461 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4462 }
4463
4464 //-------------------------------new_array-------------------------------------
4465 // helper for newarray and anewarray
4466 // The 'length' parameter is (obviously) the length of the array.
4467 // The optional arguments are for specialized use by intrinsics:
4468 // - If 'return_size_val', report the non-padded array size (sum of header size
4469 // and array body) to the caller.
4470 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4471 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4472 Node* length, // number of array elements
4473 int nargs, // number of arguments to push back for uncommon trap
4474 Node* *return_size_val,
4475 bool deoptimize_on_exception,
4476 Node* init_val) {
4477 jint layout_con = Klass::_lh_neutral_value;
4478 Node* layout_val = get_layout_helper(klass_node, layout_con);
4479 bool layout_is_con = (layout_val == nullptr);
4480
4481 if (!layout_is_con && !StressReflectiveCode &&
4482 !too_many_traps(Deoptimization::Reason_class_check)) {
4483 // This is a reflective array creation site.
4484 // Optimistically assume that it is a subtype of Object[],
4485 // so that we can fold up all the address arithmetic.
4486 layout_con = Klass::array_layout_helper(T_OBJECT);
4487 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4488 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4489 { BuildCutout unless(this, bol_lh, PROB_MAX);
4490 inc_sp(nargs);
4491 uncommon_trap(Deoptimization::Reason_class_check,
4492 Deoptimization::Action_maybe_recompile);
4493 }
4494 layout_val = nullptr;
4495 layout_is_con = true;
4496 }
4497
4498 // Generate the initial go-slow test. Make sure we do not overflow
4499 // if length is huge (near 2Gig) or negative! We do not need
4500 // exact double-words here, just a close approximation of needed
4501 // double-words. We can't add any offset or rounding bits, lest we
4502 // take a size -1 of bytes and make it positive. Use an unsigned
4503 // compare, so negative sizes look hugely positive.
4504 int fast_size_limit = FastAllocateSizeLimit;
4505 if (layout_is_con) {
4506 assert(!StressReflectiveCode, "stress mode does not use these paths");
4507 // Increase the size limit if we have exact knowledge of array type.
4508 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4509 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4510 }
4511
4512 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4513 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4514
4515 // --- Size Computation ---
4516 // array_size = round_to_heap(array_header + (length << elem_shift));
4517 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4518 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4519 // The rounding mask is strength-reduced, if possible.
4520 int round_mask = MinObjAlignmentInBytes - 1;
4521 Node* header_size = nullptr;
4522 // (T_BYTE has the weakest alignment and size restrictions...)
4523 if (layout_is_con) {
4524 int hsize = Klass::layout_helper_header_size(layout_con);
4525 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4526 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4527 if ((round_mask & ~right_n_bits(eshift)) == 0)
4528 round_mask = 0; // strength-reduce it if it goes away completely
4529 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4530 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4531 assert(header_size_min <= hsize, "generic minimum is smallest");
4532 header_size = intcon(hsize);
4533 } else {
4534 Node* hss = intcon(Klass::_lh_header_size_shift);
4535 Node* hsm = intcon(Klass::_lh_header_size_mask);
4536 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4537 header_size = _gvn.transform(new AndINode(header_size, hsm));
4538 }
4539
4540 Node* elem_shift = nullptr;
4541 if (layout_is_con) {
4542 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4543 if (eshift != 0)
4544 elem_shift = intcon(eshift);
4545 } else {
4546 // There is no need to mask or shift this value.
4547 // The semantics of LShiftINode include an implicit mask to 0x1F.
4548 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4549 elem_shift = layout_val;
4598 }
4599 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4600
4601 if (return_size_val != nullptr) {
4602 // This is the size
4603 (*return_size_val) = non_rounded_size;
4604 }
4605
4606 Node* size = non_rounded_size;
4607 if (round_mask != 0) {
4608 Node* mask1 = MakeConX(round_mask);
4609 size = _gvn.transform(new AddXNode(size, mask1));
4610 Node* mask2 = MakeConX(~round_mask);
4611 size = _gvn.transform(new AndXNode(size, mask2));
4612 }
4613 // else if round_mask == 0, the size computation is self-rounding
4614
4615 // Now generate allocation code
4616
4617 // The entire memory state is needed for slow path of the allocation
4618 // since GC and deoptimization can happen.
4619 Node *mem = reset_memory();
4620 set_all_memory(mem); // Create new memory state
4621
4622 if (initial_slow_test->is_Bool()) {
4623 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4624 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4625 }
4626
4627 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4628 const TypeOopPtr* ary_type = ary_klass->as_exact_instance_type();
4629
4630 Node* raw_init_value = nullptr;
4631 if (init_val != nullptr) {
4632 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4633 if (ary_type->is_flat()) {
4634 initial_slow_test = intcon(1);
4635 }
4636
4637 if (UseCompressedOops) {
4638 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4639 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4640 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4641 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4642 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4643 } else {
4644 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4645 }
4646 }
4647
4648 Node* valid_length_test = _gvn.intcon(1);
4649 if (ary_type->isa_aryptr()) {
4650 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4651 jint max = TypeAryPtr::max_array_length(bt);
4652 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4653 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4654 }
4655
4656 // Create the AllocateArrayNode and its result projections
4657 AllocateArrayNode* alloc
4658 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4659 control(), mem, i_o(),
4660 size, klass_node,
4661 initial_slow_test,
4662 length, valid_length_test,
4663 init_val, raw_init_value);
4664 // Cast to correct type. Note that the klass_node may be constant or not,
4665 // and in the latter case the actual array type will be inexact also.
4666 // (This happens via a non-constant argument to inline_native_newArray.)
4667 // In any case, the value of klass_node provides the desired array type.
4668 const TypeInt* length_type = _gvn.find_int_type(length);
4669 if (ary_type->isa_aryptr() && length_type != nullptr) {
4670 // Try to get a better type than POS for the size
4671 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4672 }
4673
4674 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4675
4676 array_ideal_length(alloc, ary_type, true);
4677 return javaoop;
4678 }
4679
4680 // The following "Ideal_foo" functions are placed here because they recognize
4681 // the graph shapes created by the functions immediately above.
4682
4683 //---------------------------Ideal_allocation----------------------------------
4778 void GraphKit::add_parse_predicates(int nargs) {
4779 if (ShortRunningLongLoop) {
4780 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4781 // walking up from the loop.
4782 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4783 }
4784 if (UseLoopPredicate) {
4785 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4786 if (UseProfiledLoopPredicate) {
4787 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4788 }
4789 }
4790 if (UseAutoVectorizationPredicate) {
4791 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4792 }
4793 // Loop Limit Check Predicate should be near the loop.
4794 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4795 }
4796
4797 void GraphKit::sync_kit(IdealKit& ideal) {
4798 reset_memory();
4799 set_all_memory(ideal.merged_memory());
4800 set_i_o(ideal.i_o());
4801 set_control(ideal.ctrl());
4802 }
4803
4804 void GraphKit::final_sync(IdealKit& ideal) {
4805 // Final sync IdealKit and graphKit.
4806 sync_kit(ideal);
4807 }
4808
4809 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4810 Node* len = load_array_length(load_String_value(str, set_ctrl));
4811 Node* coder = load_String_coder(str, set_ctrl);
4812 // Divide length by 2 if coder is UTF16
4813 return _gvn.transform(new RShiftINode(len, coder));
4814 }
4815
4816 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4817 int value_offset = java_lang_String::value_offset();
4818 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4819 false, nullptr, Type::Offset(0));
4820 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4821 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::BotPTR,
4822 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4823 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4824 Node* p = basic_plus_adr(str, str, value_offset);
4825 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4826 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4827 return must_be_not_null(load, true);
4828 }
4829
4830 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4831 if (!CompactStrings) {
4832 return intcon(java_lang_String::CODER_UTF16);
4833 }
4834 int coder_offset = java_lang_String::coder_offset();
4835 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4836 false, nullptr, Type::Offset(0));
4837 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4838
4839 Node* p = basic_plus_adr(str, str, coder_offset);
4840 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4841 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4842 return load;
4843 }
4844
4845 void GraphKit::store_String_value(Node* str, Node* value) {
4846 int value_offset = java_lang_String::value_offset();
4847 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4848 false, nullptr, Type::Offset(0));
4849 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4850
4851 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4852 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4853 }
4854
4855 void GraphKit::store_String_coder(Node* str, Node* value) {
4856 int coder_offset = java_lang_String::coder_offset();
4857 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4858 false, nullptr, Type::Offset(0));
4859 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4860
4861 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4862 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4863 }
4864
4865 // If input and output memory types differ, capture the whole memory to preserve
4866 // the dependency between preceding and subsequent loads/stores.
4867 // For example, the following program:
4868 // StoreB
4869 // compress_string
4870 // LoadB
4871 // has this memory graph (use->def):
4872 // LoadB -> compress_string -> CharMem
4873 // ... -> StoreB -> ByteMem
4874 // The intrinsic hides the dependency between LoadB and StoreB, causing
4875 // the load to read from memory not containing the result of the StoreB.
4876 // The correct memory graph should look like this:
4877 // LoadB -> compress_string -> MergeMem -> StoreB
4878 Node* GraphKit::capture_memory(const TypePtr*& combined_type, const TypePtr* src_type, const TypePtr* dst_type) {
4981 i_char->init_req(2, AddI(i_char, intcon(2)));
4982
4983 set_control(IfFalse(iff));
4984 set_memory(st, TypeAryPtr::BYTES);
4985 }
4986
4987 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4988 if (!field->is_constant()) {
4989 return nullptr; // Field not marked as constant.
4990 }
4991 ciInstance* holder = nullptr;
4992 if (!field->is_static()) {
4993 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4994 if (const_oop != nullptr && const_oop->is_instance()) {
4995 holder = const_oop->as_instance();
4996 }
4997 }
4998 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4999 /*is_unsigned_load=*/false);
5000 if (con_type != nullptr) {
5001 Node* con = makecon(con_type);
5002 if (field->type()->is_inlinetype()) {
5003 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
5004 } else if (con_type->is_inlinetypeptr()) {
5005 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
5006 }
5007 return con;
5008 }
5009 return nullptr;
5010 }
5011
5012 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
5013 const Type* obj_type = obj->bottom_type();
5014 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
5015 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
5016 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
5017 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
5018 obj = casted_obj;
5019 }
5020 if (!maybe_larval && sig_type->is_inlinetypeptr()) {
5021 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
5022 }
5023 return obj;
5024 }
|